Elementary Functional Analysis 0387855297, 9780387855295

This nicely written manuscript takes a gentler approach than other functional analysis graduate texts, and includes an i

480 109 3MB

English Pages [211] Year 2009

Report DMCA / Copyright

DOWNLOAD FILE

Polecaj historie

Elementary Functional Analysis
 0387855297, 9780387855295

Citation preview

Graduate Texts in Mathematics

253

Editorial Board

S. Axler K.A. Ribet

Graduate Texts in Mathematics

For other titles published in this series, go to http://www.springer.com/series/136

Barbara D. MacCluer

Elementary Functional Analysis

ABC

Barbara D. MacCluer Department of Mathematics University of Virginia P.O.Box 400137 Charlottesville VA 22904-4137 USA [email protected]

Editorial Board S. Axler

K.A. Ribet

Mathematics Department San Francisco State University San Francisco, CA 94132 USA [email protected]

Mathematics Department University of California, Berkeley Berkeley, CA 94720-3840 USA [email protected]

ISBN: 978-0-387-85528-8 DOI: 10.1007/978-0-387-85529-5

e-ISBN: 978-0-387-85529-5

Library of Congress Control Number: 2008937759 Mathematics Subject Classification ( 2000 ) : 46-01, 47-01 c Springer Science+Business Media, LLC 2009 ° All rights reserved. This work may not be translated or copied in whole or in part without the written permission of the publisher (Springer Science+Business Media, LLC, 233 Spring Street, New York, NY 10013, USA), except for brief excerpts in connection with reviews or scholarly analysis. Use in connection with any form of information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed is forbidden. The use in this publication of trade names, trademarks, service marks, and similar terms, even if they are not identified as such, is not to be taken as an expression of opinion as to whether or not they are subject to proprietary rights. Printed on acid-free paper springer.com

To Tom, Josh, and David

Preface Functional analysis arose in the early twentieth century and gradually, conquering one stronghold after another, became a nearly universal mathematical doctrine, not merely a new area of mathematics, but a new mathematical world view. Its appearance was the inevitable consequence of the evolution of all of nineteenth-century mathematics, in particular classical analysis and mathematical physics. Its original basis was formed by Cantor’s theory of sets and linear algebra. Its existence answered the question of how to state general principles of a broadly interpreted analysis in a way suitable for the most diverse situations. A.M. Vershik ([45], p. 438).

This text evolved from the content of a one semester introductory course in functional analysis that I have taught a number of times since 1996 at the University of Virginia. My students have included first and second year graduate students preparing for thesis work in analysis, algebra, or topology, graduate students in various departments in the School of Engineering and Applied Science, and several undergraduate mathematics or physics majors. After a first draft of the manuscript was completed, it was also used for an independent reading course for several undergraduates preparing for graduate school. While this book is short, comparatively speaking, it does not accomplish it aims through brevity. Arguments are generally presented in detail, and in fact I have tried to firmly keep in mind the reader who may be learning the material on his or her own without the benefit of a formal course or instructor. Since functional analysis is a huge field, I have had to make many omissions with regard to the topics I present. These choices represent, of course, my own preferences, but also my desire to start with the basics and still travel a path through some significant parts of modern functional analysis. The prerequisites for this book include undergraduate courses in real analysis, linear algebra, and basic point set topology (say, in metric spaces). A modicum of complex analysis is used in a few examples and exercises, and in the proofs of a few results in Chapter 5; in a pinch it is not an essential prerequisite for a student willing to bypass those parts (or take them on faith). With respect to real analysis a good undergraduate level course is essential. Beyond this some familiarity with measure theory and the Lebesgue integral is desirable, but not essential. Save for the last chapter, most of the use of measure theory and Lebesgue integration occurs in limited ways—primarily in examples. An Appendix provides a summary and expository discussion of all that is needed here. I encourage any prospective reader who may feel shaky with these desirable but not essential prerequisites not to be daunted by them. On the basis of my experiences teaching this material I have found that students with no prior exposure to complex analysis or measure theory and Lebesgue integration can nevertheless have a successful experience with the topics presented here.

vii

viii

Preface

I have woven a certain amount of historical commentary into the text; this reflects my belief that some understanding of the historical development of any field in mathematics both deepens and enlivens one’s appreciation of the subject. The history of functional analysis is filled with interesting characters, many of whom lived and worked during turbulent times in the twentieth century. Each chapter concludes with an extensive collection of exercises. The purpose of the exercises is to enable the reader to become comfortable with the ideas in the text; to make them his or her own. While most are therefore closely tied to the material being discussed, an occasional exercise is intended to provide an initial step or steps towards a topic not discussed in the text, or to point the way for further exploration. In any case, all are intended to be eminently doable by a student and when advisable are accompanied by a hint. I would like to express my great appreciation to several friends and colleagues who provided advice and encouragement during the writing of this book. Sheldon Axler, Tom Goebeler, Christopher Hammond, and Bill Ross read substantial portions of the manuscript and provided many helpful comments, as well as suggestions for exercises. Larry Thomas gave useful feedback on the Appendix. Mark Spencer at Springer provided valuable editorial assistance. Julie Riddleberger helped with the illustrations, and patiently answered many TEX questions. I thank Tom Kriete for his enthusiastic support and encouragement throughout all stages of this work. And finally I thank the students in the Functional Analysis course I taught at the University of Virginia in each of the last several years. It was their enthusiastic response to this course that initially got me thinking about writing a functional analysis text, and helped me refine my ideas of what this text should look like. Charlottesville, Virginia

Barbara D. MacCluer July 2008

Contents

Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii 1

Hilbert Space Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.1 Normed Linear Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2 Orthogonality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.3 Hilbert Space Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.4 Linear Functionals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.5 Orthonormal Bases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1 2 10 12 15 19 23

2

Operator Theory Basics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.1 Bounded Linear Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2 Adjoints of Hilbert Space Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.3 Adjoints of Banach Space Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

31 31 34 41 43

3

The Big Three . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.1 The Hahn–Banach Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2 Principle of Uniform Boundedness . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.3 Open Mapping and Closed Graph Theorems . . . . . . . . . . . . . . . . . . . . 3.4 Quotient Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.5 Banach and the Scottish Caf´e . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

49 50 55 61 66 67 68

4

Compact Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77 4.1 Finite-Dimensional Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77 4.2 Compact Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80 4.3 A Preliminary Spectral Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87 4.4 The Invariant Subspace Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94 4.5 Introduction to the Spectrum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96 4.6 The Fredholm Alternative . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99 4.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101 ix

x

Contents

5

Banach and C∗ -Algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107 5.1 First Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108 5.2 Results on Spectra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110 5.3 Ideals and Homomorphisms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120 5.4 Commutative Banach Algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124 5.5 Weak Topologies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127 5.6 The Gelfand Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132 5.7 The Continuous Functional Calculus . . . . . . . . . . . . . . . . . . . . . . . . . . 140 5.8 Fredholm Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143 5.9 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146

6

The Spectral Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157 6.1 Normal Operators Are Multiplication Operators . . . . . . . . . . . . . . . . . 157 6.2 Spectral Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165 6.3 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183

Appendix A: Real Analysis Topics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187 A.1 Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187 A.2 Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190 A.3 L p Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196 A.4 The Stone–Weierstrass Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197 A.5 Positive Linear Functionals on C(X) . . . . . . . . . . . . . . . . . . . . . . . . . . . 198 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201 Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203

Chapter 1

Hilbert Space Preliminaries

It seems to me not useless to indicate interest in a study of sets composed of functions.... J. Hadamard, International Congress of Mathematicians, Z˜urich, 1897.

Functional analysis developed in the late nineteenth and early twentieth centuries, during a period in which there was a general interest in abstraction, axiomatization, and unification across all fields of mathematics. This unification meant that objects that behaved according to a common set of rules were viewed as “the same,” even if they consisted of rather different elements. A core idea in functional analysis is to treat functions as “points” or “elements” in some sort of abstract space, so that instead of working with individual functions (the tradition in classical analysis), we deal with functions as points in a space endowed with some kind of overall structure. The structure of the space itself is emphasized over properties of individual elements in the space. This viewpoint, accompanied by an axiomatization of the new spaces to be considered, was an integral step in the process of transferring familiar concepts in finite-dimensional Euclidean space to (typically infinite-dimensional) “function spaces.” While important contributions to the beginnings of functional analysis were made by individuals of various nationalities, the most readily identifiable schools of work in the early history of the subject were in France, Italy, and Germany. In France, one of the notable contributors to the initial development of functional analysis was Maurice Fr´echet, whose 1906 doctoral dissertation is a landmark paper in the subject. In this work, which was extremely influential in both functional analysis and point set topology, Fr´echet began the study of abstract spaces of functions. In particular, he defined the notion of a metric space (which he called “(E)” spaces, from the French “´ecart” meaning distance), and included a discussion of examples of metric spaces where the points in the space were functions. In Fr´echet’s work one can clearly see the influence of his advisor Jacques Hadamard. In an address to the International Congress of Mathematicians in 1897, Hadamard proposed a study of what would now be termed set-theoretic topology. A quote from this address introduces this chapter; his student Fr´echet took up the challenge put forth there. In this chapter we describe the basic kinds of spaces which will interest us, with a particular emphasis on Hilbert spaces, which are rich in geometric structure. In simplest terms, the idea behind a Hilbert space is to generalize the familiar Euclidean

B.D. MacCluer, Elementary Functional Analysis, DOI 10.1007/978-0-387-85529-5 1, c Springer Science+Business Media, LLC 2009 

1

2

1 Hilbert Space Preliminaries

spaces Rn or Cn , preserving as much as possible the geometric results in these finitedimensional settings.

1.1 Normed Linear Spaces A modern-seeming, axiomatic, definition of vector spaces goes back to the Italian mathematician Giuseppe Peano, in 1888. A vector space is an algebraic object; to introduce such analytic notions as convergence or continuity in a vector space we must provide our vector space with additional structure. This brings us to the concept of a normed linear space, which is a vector space with a norm. Definition 1.1. Let X be a vector space over either the scalar field R of real numbers or the scalar field C of complex numbers. Suppose we have a function  ·  : X → [0, ∞) such that (1) (2) (3)

x = 0 if and only if x = 0, x + y ≤ x + y for all x, y ∈ X, and α x = |α |x for all scalars α and vectors x.

We call (X,  · ) a normed linear space. Property (2) is called the triangle inequality, and property (3) is referred to as homogeneity. The reverse triangle inequality, x + y ≥ |x − y| follows easily from (2); see Exercise 1.1. We give some examples of normed linear spaces. In these examples we won’t give the details of the verification that the norm satisfies these defining properties. This verification is straightforward in some cases, while in others it may already be known to the reader or will be outlined in an exercise. Example 1.2. Let X = Cn ≡ {(z1 , z2 , . . . , zn ) : z j ∈ C} with n

(z1 , z2 , . . . , zn ) = ( ∑ |z j |2 ) 2 ; 1

j=1

this is called the Euclidean norm. The Euclidean space Rn is similarly defined; in this case we restrict to real scalars. Example 1.3. Let X = Cn with (z1 , z2 , . . . , zn ) = max{|z j | : 1 ≤ j ≤ n}. Example 1.4. Let Y = [0, 1], or more generally any compact Hausdorff space, and let C(Y ) be the vector space of continuous, complex-valued functions on Y , under pointwise addition and scalar multiplication. Define a norm on C(Y ) by  f  = max{| f (y)| : y ∈ Y }. This (specifically C[a, b], endowed with the metric which defines the distance between functions f and g to be maxa≤x≤b | f (x) − g(x)|), was one of the important examples that Fr´echet put forth in his 1906 dissertation.

1.1 Normed Linear Spaces

3

Example 1.5. Choose a value of p ≥ 1, and let  p =  p (N) denote the set of all sequences {an }∞ n=1 of complex numbers (indexed by the positive integers N) for p < ∞. In our notation for a sequence we will often abbreviate {a }∞ |a | which ∑∞ n n=1 1 n p by {an }∞ 1 or even just {an }. Define the norm of {an } ∈  by 

1/p



∑ |an |

{an } p ≡

p

.

1

We can include the choice p = ∞ by modifying this definition in the expected way: ∞ = {{an }∞ 1 : sup |an | < ∞} n

and {an }∞ = sup |an |. n

For p = 1 and p = ∞ the triangle inequality is easily verified; for 1 < p < ∞ it goes by the name of Minkowski’s inequality, in honor of Hermann Minkowski who first studied the analogue of this  p -norm on the space Rn . Example 1.6. We can generalize the last example as follows. Consider a positive measure space (Y, M, µ ), where Y is a set, M is a σ -algebra of subsets of Y , and µ is a positive measure. Choose 1 ≤ p < ∞, and denote by L p (Y, µ ) the collection of all equivalence classes of M-measurable functions on Y with  Y

| f | p d µ < ∞,

normed by  f p =



1 | f | dµ p

Y

p

(the integral in this definition is the Lebesgue integral). Minkowski’s inequality (for integrals) provides the proof that the norm satisfies the triangle inequality. We also define L∞ (X, µ ) to be all equivalence classes of essentially bounded measurable functions, normed by  f ∞ = ess sup | f |, the essential supremum of f . Of particular interest to us will be the space L p [0, 1] = L p ([0, 1], dx) with respect to Lebesgue measure dx on the real line. For the reader unfamiliar with the concepts in the preceding example, the Appendix provides a summary of the relevant definitions and results from real analysis. The use of the Lebesgue integral in the definition of the L p spaces is important, and in writing a history of functional analysis, Jean Dieudonn´e [10] states ...it is likely that progress in Functional Analysis might have been appreciably slowed down if the invention of the Lebesgue integral had not appeared, by a happy coincidence, exactly at the beginning of Hilbert’s work...(pp. 119–120).

replacing, what Dieudonn´e calls “the horrible and useless so-called Riemann integral.” In Example 1.6, the particular choice Y = N and µ = counting measure on

4

1 Hilbert Space Preliminaries

the subsets of N gives the space  p of Example 1.5; see Sections A.2 and A.3 in the Appendix for more details. Example 1.7. Fix a sequence {β (n)}∞ n=0 of positive numbers with β (0) = 1 and lim β (n)1/n ≥ 1.

(1.1)

n→∞

The reason for this last restriction will be made clear shortly, but for right now notice that defining β (n) = (n + 1)a for some fixed real number a will give an allowable choice. Define the weighted sequence space 2β to consist of all sequences {an }∞ 0 with ∞

∑ |an |2 β (n)2 < ∞,

n=0

where the norm of

{an }∞ 0

is defined to be 



1/2

∑ |an |2 β (n)2

.

n=0

From one perspective these weighted sequence spaces can be thought of simply as L2 (X, M, µ ) for X = N0 ≡ {0} ∪ N, M the collection of all subsets of N0 , and µ the measure that assigns to each point n of N0 the mass β (n)2 , so that we have a special case of the example discussed in Example 1.6. In particular, the general version of Minkowski’s inequality gives the triangle inequality in 2β . (See Exercise 1.6 for a more elementary approach.) The requirement in Equation (1.1) allows us to offer a second perspective on the spaces 2β , and the interplay between the two perspectives endows these examples 2 with a particular richness. Associate to a sequence {an }∞ 0 in β the power series ∞ n ∑n=0 an z . The radius of convergence of this series is at least one (see Exercise 1.9), and thus the series converges to an analytic function on the unit disk D = {z ∈ C : |z| < 1}. This suggests that we may want to identify 2β , a space of sequences, with the vector space ∞



0

0

{ f = ∑ an zn analytic in D : ∑ |an |2 β (n)2 < ∞}. In the latter guise, the space is referred to as a weighted Hardy space and denoted H 2 (β ); the case β (n) = 1 for all n gives the Hardy space H 2 . In the next chapter we will have the language needed to make precise the properties of this identification, ∞ n but for the moment we simply observe that the map sending {an }∞ 0 to f = ∑0 an z 2 is one-to-one (by uniqueness of power series) and onto H (β ) by definition, and we will regard H 2 (β ) as normed so that this mapping preserves norms. Example 1.8. Let Ω be a nonempty open set in C. Denote the collection of all bounded analytic functions on Ω by H ∞ (Ω ), and introduce a norm on H ∞ (Ω ) by  f  = sup{| f (z)| : z ∈ Ω }.

1.1 Normed Linear Spaces

5

The norms in Examples 1.3, 1.4, 1.8, and the ∞ norm in Example 1.5, are all referred to as the “supremum norm,” and when needed for clarity will be written as  · ∞ . The L∞ norm in Example 1.6 is called the essential supremum norm; it is also written  · ∞ . Definition 1.9. A metric space is a set X with a function d(·, ·) : X × X → [0, ∞) satisfying, for x, y, and z in X, (1) (2) (3)

d(x, y) = 0 if and only if x = y, d(x, y) = d(y, x), and d(x, y) + d(y, z) ≥ d(x, z).

The third property is referred to as the triangle inequality. On any metric space (X, d) there is an associated topology. The open balls are the sets of the form B(a, r) ≡ {x : d(x, a) < r}, where r > 0. Every open set is the union of some collection of open balls. It is easy to see that if X is a normed linear space, we may define a metric on X by defining d(x, y) = x − y. With the metric topology in place on X, continuity of certain basic mappings can be addressed. For example, it is easy to check that the function  ·  : X → [0, ∞) is continuous; see Exercise 1.8 for this and other elementary results. About eight years after Fr´echet’s seminal work in 1906, Felix Hausdorff wrote a text that presented a thoroughly modern definition of metric space and defined the fundamental idea of a Cauchy sequence, which we recall next. Definition 1.10. Let X be a metric space. A sequence {xn } in X is said to be a Cauchy sequence if it has the following property: Given any ε > 0 there exists N such that if n, m ≥ N, then d(xn , xm ) < ε . Definition 1.11. A metric space is said to be complete if every Cauchy sequence in X converges in X. Definition 1.12. Let X be a normed linear space. If X is complete in the metric d defined from the norm by d(x, y) = x − y, we call X a Banach space. All of the above examples of normed linear spaces are Banach spaces. We will not stop to prove this now, but we do make a couple of observations. The statement that the space L p (Y, µ ) is complete for any 1 ≤ p ≤ ∞ and any positive measure space (Y, µ ) goes by the name of the Riesz–Fischer theorem. In its full generality it is a deep result of real analysis (see also the discussion in Section 1.5 below and in Section A.3 of the Appendix). Notice that this general class of examples includes the  p spaces and weighted sequences spaces as special cases (see Exercise 1.6 for a more elementary approach), as well as the finite-dimensional spaces in Examples 1.2 and 1.3. In Exercise 1.2 the reader is asked to provide a proof of completeness for the spaces in Example 1.4, and a similar argument can be used for the space H ∞ (Ω ) of Example 1.8. You can get an example of a normed linear space which is not a Banach space by taking a nonclosed subspace of a Banach space; see

6

1 Hilbert Space Preliminaries

for example Exercise 1.3. (A subspace of a vector space V is a subset of V which is itself a vector space under the same addition and scalar multiplication operations.) Banach spaces are named in honor of the Polish mathematician Stefan Banach, a dominating figure in the birth of functional analysis, who wrote a fundamentally important book called Op´erations Lin´eaires in 1932. In this book (which had its beginnings in Banach’s 1920 doctoral thesis) many of the properties of complete normed linear spaces are developed. Banach calls these spaces “spaces of type (B),” perhaps in the hope they would eventually be known as “Banach spaces”1 This is precisely what happened, with the terminology “Banach space” making its formal appearance in Fr´echet’s text Les Espaces Abstraits [13]. Hugo Steinhaus, Banach’s teacher and collaborator, writes in a 1963 memoir of Banach that Banach’s axiomatic definition of a complete normed linear space provided precisely the right level of generality; broad enough to encompass a wide variety of natural examples, but not so general as to permit only uninteresting theorems: His foreign competitors in the theory of linear operations either dealt with spaces that were too general, and that is why they either obtained only trivial results, or assumed too much about those spaces, which restricted the extent of the applications to a few and artificial examples — Banach’s genius reveals itself in finding the golden mean. This ability of hitting the mark proves that Banach was born a high class mathematician ([44], p. 12).

In fact, a few months after Banach set down the axioms for a normed linear space, the American Norbert Wiener independently gave nearly the same definition, and for a short while the terminology “Banach–Wiener spaces” was used. However, as Wiener’s interest in the area did not continue, these spaces, in Wiener’s words, became “quite justly named after Banach alone ([46], p. 60).” Hilbert spaces, which we turn to now, are Banach spaces with some additional structure, coming from the presence of an inner product. Definition 1.13. Let X be a vector space over C. An inner product is a map ·, · : X × X → C satisfying, for x, y, and z in X and scalars α ∈ C, (1) (2) (3) (4)

x, y = y, x for all x, y in X, x, x ≥ 0, with x, x = 0 (if and) only if x = 0, x + y, z = x, z + y, z , and α x, y = α x, y .

Some comments on this definition are in order. The bar in (1) denotes complex conjugation. Property (2) is referred to as “positive-definiteness,” and the adjective “Hermitian” is used for property (1). The parenthetical “if” statement in (2) need not be included in the definition, as it follows from the other parts since 0, 0 = 2 · 0, 0 = 2 0, 0 . An inner product is linear in the first slot and conjugate linear in 1 Though this interpretation of Banach’s choice of notation is widely repeated, V.D. Milman, in writing about Banach, says, “In his book...Banach denotes operators by the letter A. These were the initial objects of study, and the complete normed spaces on which they operated were denoted by the Latin letter B. That was natural, and there is no indication that he was ‘hinting’ at his own name by using that letter” ([32], p. 228).

1.1 Normed Linear Spaces

7

the second ( x, α y+z = α x, y + x, z ), so the defining properties are encapsulated by saying that an inner product is a Hermitian, positive definite, sesquilinear form (sesquilinear from the Latin for “1 12 ” linear). The reader is cautioned that some authors (in physics, for example) define the inner product to be linear in the second slot, and conjugate linear in the first. A standard example is to define an inner product on L2 (X, µ ) for a positive measure space (X, µ ) by  f , g =

X

f g dµ.

This general framework includes, as special cases, the example Cn with (z1 , z2 , . . . , zn ), (w1 , w2 , . . . , wn ) =

n

∑ z jw j,

j=1

the example 2 of all square summable sequences with ∞

(z1 , z2 , . . .), (w1 , w2 , . . .) =

∑ z jw j,

j=1

and the weighted analogues 2β with (z0 , z1 , . . .), (w0 , w1 , . . .) =



∑ z j w j β ( j)2 .

j=0

The first two are obtained by taking X to be, respectively, {1, 2, . . . , n} or N, with µ equal to counting measure. In the case of weighted sequence spaces, X = N0 and µ assigns mass β (n)2 to the set {n}. Any inner product satisfies an important inequality, called the Cauchy–Schwarz inequality, which we describe next. Proposition 1.14. If ·, · is an inner product on a vector space X, then for all x and y in X we have | x, y |2 ≤ x, x y, y . In this general form, the Cauchy–Schwarz inequality is due to John von Neumann (1930), who is often credited with the “axiomatization” of Hilbert spaces (defined below). Earlier versions of Proposition 1.14, for specific settings, go back to Cauchy, Bunyakowsky, and Schwarz, and the Cauchy–Schwarz inequality is sometimes referred to as the Cauchy–Bunyakowsky–Schwarz inequality. One particularly simple proof of Proposition 1.14 is outlined in Exercise 1.7. As an important application of Proposition 1.14, we show next how any inner product defines a norm. Proposition 1.15. If ·, · is an inner product on a vector space X, then 1

x ≡ x, x 2

8

1 Hilbert Space Preliminaries

is a norm on X. Proof. We will check the triangle inequality, and leave the verification of the other norm properties to the reader. Using the linearity of the inner product we have x + y2 = x + y, x + y = x, x + y, x + x, y + y, y

= x2 + 2Re x, y + y2 ≤ x2 + 2| x, y | + y2 ≤ x2 + 2xy + y2 = (x + y)2 where Re z denotes the real part of a complex number z, and we have used the Cauchy–Schwarz inequality in the penultimate step. Definition 1.16. A (complex) Hilbert space H is a vector space over C with an inner product such that H is complete in the metric 1

d(x, y) = x − y = x − y, x − y 2 . Any space L2 (X, µ ) as described above is thus an example of a Hilbert space, since we have already observed that L2 (X, µ ) is a Banach space under the norm  1 1  f 2 = ( X | f |2 d µ ) 2 which we recognize as f , f 2 . There are various anecdotes, of dubious validity, about David Hilbert and the terminology “Hilbert space.” Steve Krantz, writing in Mathematical Apocrypha [27] says It is said that, late in his life, Hilbert was reading a paper and got stuck at one point. He went to his colleague in the office next door and queried, “What is a Hilbert space?” (p. 89)

Another version is given by Laurence Young [47]: When Weyl presented a proof of the Riesz–Fischer theorem in a G¨ottingen colloquium, Hilbert went up to the speaker afterward to say,“Weyl, you must just tell me one thing, whatever is a Hilbert space?” (p. 312)

Next we will look at an important example of a Hilbert space where the vectors are certain analytic functions on the unit disk D = {z ∈ C : |z| < 1}. This example, which uses a few basic results from complex analysis, will prove to be particularly illuminating of several of the fundamental Hilbert space notions. Example 1.17. The Bergman space La2 (D) is the vector space, under pointwise operations, of all analytic functions f on D for which  D

| f (z)|2

dA < ∞, π

where dA denotes two-dimensional Lebesgue measure (so that dA/π is “normalized area measure” on the unit disk). Of course, every function in La2 (D) is (a representative of) an element of the Hilbert space L2 (D, dA/π ); we give La2 (D) the inner product it inherits from L2 (D):

1.1 Normed Linear Spaces

9

f , g =



f (z)g(z)

D

dA . π

Our first goal is to check that the Bergman space is a Hilbert space. How much work must we do? Since we already know that L2 (D, dA/π ) is a Hilbert space, it will suffice to verify that La2 (D) is a closed subspace of L2 (D, dA/π ). Of course, “closed” here refers to the topology on L2 (D, dA/π ); this is the metric topology induced by the norm. To this end, we need an area mean-value property for analytic functions. Proposition 1.18. If f is a analytic function in some closed disk B(a, R), then f (a) =

1 π R2



f dA. B(a,R)

Proof. As a consequence of Cauchy’s integral formula we have the mean value property  1 2π f (a) = f (a + reiθ )d θ 2π 0 for all 0 < r < R. Multiplying by r and integrating with respect to r we have  R

r f (a)dr = 0

 R  2π 0

0

or equivalently f (a)

f (a + reiθ )r

R2 1 = 2 2π

dθ dr 2π



f dA B(a,R)



as desired.

From this we get a corollary that gives an upper bound on the value of a function in the Bergman space at a point w ∈ D in terms of the norm of f and the distance from w to ∂ D, the unit circle. Corollary 1.19. Fix w ∈ D. For every f ∈ La2 (D) we have | f (w)| ≤

1  f La2 (D) . 1 − |w|

Proof. Let 0 < r < 1 − |w| so that the closed disk B(w, r) is contained in D. Using Proposition 1.18 and H¨older’s inequality we have    1    | f (w)| =  2 f dA π r B(w,r)  1 ≤ 2 | f | dA π r B(w,r) 1/2  1/2  1 ≤ 2 1 dA | f |2 dA πr B(w,r) B(w,r)

10

1 Hilbert Space Preliminaries

 1/2 1 √ 2√ 2 dA π r π | f | π r2 π D 1 =  f La2 (D) . r ≤

This calculation holds for any r < 1 − |w|, so letting r increase to 1 − |w| yields the desired conclusion. To show that the Bergman space is a Hilbert space, we will use, in addition to Corollary 1.19, a result from real analysis that says if a sequence { fn } in L2 (D, dA/π ) converges in the L2 -norm to a limit f , then some subsequence { fnk } converges pointwise almost everywhere (dA/π ) to f ; see, for example, p. 74 in [40]. Theorem 1.20. The Bergman space La2 (D) is a Hilbert space. Proof. As we have discussed, we need only show that La2 (D) is a closed subspace of L2 (D, dA/π ). That La2 (D) is a subspace is immediate. To see that it is closed, suppose we have a sequence { fn } of functions in La2 (D) with fn → f in L2 (D, dA/π ). Our task is to show that f must be in La2 (D); that is, that the limit function f is analytic (or more precisely, has an analytic representative). On the one hand, from the remark preceding the theorem, we know that convergence of fn to f in the norm of L2 (D, dA/π ) implies some subsequence { fnk } converges pointwise almost everywhere (dA/π ) to f . On the other hand, by Corollary 1.19 we have, for any closed disk B(0, r) ⊂ D and all z in this closed disk, | fn (z) − fm (z)| ≤

1  fn − fm La2 (D) . 1−r

This says that the sequence { fn } is uniformly Cauchy on B(0, r) and, by Morera’s theorem from complex analysis, fn converges uniformly on B(0, r) to an analytic function g on B(0, r). This holds for all r < 1 and thus our pointwise limit f must agree almost everywhere with an analytic function, i.e., we may choose f to be analytic in L2 (D, dA/π ). This is precisely the desired conclusion that f is in the Bergman space La2 (D). There are L p , p = 2, versions of the Bergman space; see Exercise 1.11 for the definition and some basic properties.

1.2 Orthogonality A Banach space is a complete normed linear space and a Hilbert space is a complete inner product space. The presence of an inner product permits the all-important geometric notion of orthogonality, which says in turn that Hilbert spaces behave in many ways as generalizations of finite-dimensional Euclidean space, where one can talk about angles and projections, for example.

1.2 Orthogonality

11

Definition 1.21. Given vectors f , g in a Hilbert space H , we say that f is orthogonal to g, written f ⊥ g, if f , g = 0. For sets A and B in H we write A ⊥ B if f , g = 0 for all f ∈ A and g ∈ B. Finally, A⊥ is the set of all vectors f ∈ H such that f ⊥ g for all g in A; for any set A this is always a subspace of H , moreover since A⊥ = ∩a∈A {a}⊥ , A⊥ is a closed subspace by continuity of the inner product (see Exercise 1.8). It should be clear that A ∩ A⊥ = {0}. (Why?) Some authors use the terminology “linear manifold” for a linear subspace that is not necessarily closed, and reserve the term “subspace” for a closed linear manifold. We will not do so, but instead use the adjective “closed” when it applies. An example of a subspace which is not closed is the set of all sequences in 2 with finitely many nonzero terms. The next result, aptly called the Pythagorean theorem, is easily verified by writing the norm in terms of the inner product and expanding. The details are left to the reader. Proposition 1.22. If f1 , f2 , . . . , fn are pairwise orthogonal vectors in a Hilbert space, then  f1 + f2 + · · · + fn 2 =  f1 2 +  f2 2 + · · · +  fn 2 . In general, for any vectors f and g in a Hilbert space we have  f + g2 =  f 2 + 2Re f , g + g2 and  f − g2 =  f 2 − 2Re f , g + g2 . The parallelogram equality is then obtained:  f + g2 +  f − g2 = 2 f 2 + 2g2 . Its name comes from picturing the relationship for vectors in, say, R2 ; see Figure 1.1.

 

 

f +g

f −g

g

 

f

-

 

 



FIGURE 1.1: The parallelogram equality In any inner product space, the inner product can be recovered from the norm:

12

1 Hilbert Space Preliminaries

f , g =

1  f + g2 −  f − g2 + i f + ig2 − i f − ig2 . 4

(1.2)

This is called the polarization identity, and it is verified by a straightforward calculation. It can be written, and perhaps more easily remembered, as f , g =

1 3 k ∑ i  f + ik g2 . 4 k=0

Perhaps surprisingly, given a normed linear space in which the parallelogram equality holds, there is an inner product that gives the norm. See Exercise 1.14 for an outline of how to show this result, which is due to P. Jordan and J. von Neumann. This exercise gives the best known of many (hundreds!) of ways of characterizing those normed linear spaces that are in fact inner product spaces. For much more on this subject, the reader is referred to [1].

1.3 Hilbert Space Geometry A convex set in a vector space V is a subset S of V with the property that whenever a, b are in S, so is ta + (1 − t)b for any 0 ≤ t ≤ 1. Clearly every subspace is convex, every ball in a normed linear space is convex, and any translate x + S ≡ {x + s : s ∈ S} of a convex set S is convex. The next result, which we will refer to as the nearest point property, is a key step in obtaining our main theorem on Hilbert space geometry. Proposition 1.23 (Nearest Point Property). Every nonempty, closed convex set K in a Hilbert space H contains a unique element of smallest norm. Moreover, given any h ∈ H , there is a unique k0 in K such that h − k0  = dist(h, K) ≡ inf{h − k : k ∈ K}. Proof. We begin with a proof of the first statement. The parallelogram equality says that for any vectors x, y in H ,



2

x − y 2 1





= (x2 + y2 ) − x + y .

2

2 2

If d ≡ inf{y : y ∈ K}, then we may find a sequence of vectors {xn } in K with xn  → d. Thus for any n, m we have



2

xn − xm 2 1





= (xn 2 + xm 2 ) − xn + xm ,

2

2 2

where, by convexity, 12 xn + 12 xm is in K, so that

1.3 Hilbert Space Geometry

13



xn + xm 2 2



2 ≥d . Thus we have 0 ≤ xn − xm 2 ≤ 2(xn 2 + xm 2 ) − 4d 2 . This tells us that {xn } is a Cauchy sequence, and by completeness it must converge to some x ∈ H . Since K is closed, x ∈ K. Continuity of the norm says that xn  → x, so x = d. This gives us the existence part of the first statement. For uniqueness, suppose z = x = d for some z in K. Consider 12 x + 12 z ∈ K. Since we must have



x+z



2 ≥d the parallelogram equality again says



2

2



x − z 2 1







= (x2 + z2 ) − x + z = d 2 − x + z ≤ 0,

2





2 2 2

which forces x = z. This completes the proof of the first statement. The second statement is obtained by translation. To find the unique point in K closest to a given h in H , first find the unique point x in the convex set K − h of minimal norm. Its translate x + h is the desired point. The arguments used in this proof are basically those of the Hungarian mathematician Frederic Riesz, another important contributor during the early period of functional analysis. We will continue to see his name attached to quite a few of the results discussed in this book. The nearest point property is quite rigid—it fails to be true if we omit either the requirement that K be closed or convex, or change “Hilbert space” to “Banach space” in the statement. The interested reader can provide examples to illustrate this. We will get a lot of mileage out of the next result, called the projection theorem, whose proof uses the nearest point property. Our presentation follows that of [40]. Theorem 1.24 (Projection Theorem). Let M be a closed subspace of a Hilbert space H . There is a unique pair of mappings P : H → M and Q : H → M ⊥ such that x = Px + Qx for all x ∈ H . Furthermore, P and Q have the following additional properties: (a) (b) (c) (d) (e) (f)

x ∈ M =⇒ Px = x and Qx = 0. x ∈ M ⊥ =⇒ Px = 0 and Qx = x. Px is the closest vector in M to x. Qx is the closest vector in M ⊥ to x. Px2 + Qx2 = x2 for all x. P and Q are linear maps.

Proof. First we define P as follows: For every x ∈ H , let Px be the unique closest point to x in the (closed convex) set M; here we are using the nearest point property

14

1 Hilbert Space Preliminaries

of Proposition 1.23. Uniqueness says that P is well-defined. Moreover, P : H → M, and if x ∈ M, Px = x. Define Qx = x − Px so that Q is uniquely defined on H , Px + Qx = x for all x, and if x ∈ M, Qx = 0. We next show that Qx ∈ M ⊥ for all x. It suffices to show that x − Px, m = 0 for all m ∈ M. Clearly it is enough to check this for unit vectors in M. Fix m ∈ M, m = 1. Consider Px + α m ∈ M for α any complex number. Since Px is the closest point to x in M x − (Px + α m) ≥ x − Px. Writing z = x − Px we have z2 + |α |2 m2 − α m, z − z, α m ≥ z2 . This is true for all α complex, and choosing α = z, m we see that α = 0. This says 0 = α = z, m = x − Px, m , which verifies the claim and proves that Qx ∈ M ⊥ as desired. If x ∈ M ⊥ , then x − Qx ∈ M ⊥ , since M ⊥ is a subspace. But also x − Qx = Px ∈ M, and M ∩ M ⊥ = {0}, so x ∈ M ⊥ implies x = Qx and Px = 0. Furthermore, for any x∈H , x2 = Px + Qx, Px + Qx = Px2 + Qx2 , since Qx ∈ M ⊥ and Px ∈ M. Next we check the linearity of the maps P and Q. Let x, y be arbitrary vectors in H . We want to show P(x + y) = Px + Py and similarly for Q. Since x = Px + Qx, y = Py + Qy and x + y = P(x + y) + Q(x + y) we see that Px + Qx + Py + Qy = P(x + y) + Q(x + y) so that Qx + Qy − Q(x + y) = P(x + y) − Px − Py. The vector on the left side of the last line lies in M ⊥ and the vector on the right side in M; this forces both to be 0, and we have our desired conclusions. The statements P(α x) = α Px and Q(α x) = α Qx are proved similarly. There are two remaining parts to the proof. We must show that Qx is the closest vector in M ⊥ to x, and verify the uniqueness statement for P and Q. The first of these goes as follows: Let x ∈ H , and suppose y ∈ M ⊥ . Since Qx ∈ M ⊥ for all x, x − y2 = Px + Qx − y2 = Px2 + Qx − y2 ; this is clearly minimized if y = Qx. So Qx is the closest vector to x in M ⊥ . The uniqueness of P and Q with the specified properties is easy: if P, P : H → M and Q, Q : H → M ⊥ with Px + Qx = x = P x + Q x for all x then Px − P x = Q x − Qx; the common value must be 0 since M ∩ M ⊥ = {0}. Figure 1.2 illustrates the projections P and Q from the projection theorem.

1.4 Linear Functionals

15

M⊥

6

 

 

6



Qx



x

: 



Px M



 

 

FIGURE 1.2: The projections P and Q The linear maps P and Q in the projection theorem are called the orthogonal projections of H onto M and M ⊥ , respectively. The notation H = M ⊕ M ⊥ is commonly used to encapsulate the statement of the projection theorem. We end this section with a simple, but useful, corollary of the projection theorem, whose proof is left to the reader. Corollary 1.25. If M is a closed, proper, subspace of H , then there exists a nonzero vector y in H with y ⊥ M. As a consequence of this corollary, note that one can show that a closed subspace M is all of H by showing that there is no nonzero vector y in H with y ⊥ M.

1.4 Linear Functionals Definition 1.26. If X is a normed linear space over C, a linear functional on X is a map Λ : X → C satisfying Λ (α x + β y) = αΛ (x) + βΛ (y) for all vectors x and y in X and all scalars α and β . Hadamard in 1903, and his student Fr´echet in 1904–05, began to investigate the continuous linear functionals on various function spaces. Hadamard, for example, described the linear functionals on C[a, b] as having the form

Λ ( f ) = lim

n→∞



f (x)Φn (x)dx

for a sequence of continuous functions Φn and, in a letter to Fr´echet in 1904, proposed the term “functional” for these “functions of functions.” When the function space under investigation was a Hilbert space, work done independently by Fr´echet and Riesz gave a particularly pleasant and important characterization of these linear functionals, as we will soon see. We begin with a definition.

16

1 Hilbert Space Preliminaries

Definition 1.27. A bounded linear functional on a normed linear space X is a linear functional Λ : X → C for which there exists a finite constant C satisfying |Λ (x)| ≤ Cx for all x ∈ X. Anticipating notation somewhat, we write Λ  ≡ sup{|Λ (x)| : x ≤ 1} and refer to this as the norm of Λ ; it is easy to check that when we give linear structure to the collection of all bounded linear functionals on any normed linear space by defining the vector operators of addition and scalar multiplication pointwise,  ·  is indeed a norm on this linear space. More will be said about structure on this space later. Exercise 1.16 gives several equivalent formulations of Λ , which we will use without comment in what follows. An easy, but fundamental, observation is that bounded linear functionals are precisely those linear functionals which are continuous. Proposition 1.28. If X is a normed linear space, and Λ : X → C is a linear functional, then the following are equivalent: (a) (b) (c)

Λ is continuous. Λ is continuous at 0. Λ is bounded.

Proof. The implication (a)⇒(b) is trivial, so we look first at (b)⇒(c). Since Λ (0) = 0 (why?), continuity of Λ at 0 means that given ε > 0 we may find δ > 0 such that if x ≤ δ , then |Λ (x)| < ε . Choose such a δ to correspond to ε = 1. Given x = 0, linearity tells us that     x xδ  x ≤ |Λ (x)| = Λ · , δ x  δ which gives the boundedness of Λ with Λ  ≤ 1/δ . The result of Exercise 1.16 is used here. The proof of (c)⇒(a) is left to the reader. In discussing continuity, it is helpful to recall that any map f : X → Y , where X is a metric space but Y need only be a topological space, is continuous if and only if given any sequence {xn } in X that converges to a point x0 in X, the sequence { f (xn )} converges to f (x0 ). The “if” direction of this statement need not be true if X is not a metric space; this issue will be discussed further in Section 5.5. As a consequence of a somewhat more general result that we will prove later, the set of all bounded linear functionals on a normed linear space X is itself a Banach space, under pointwise operations and using the norm just defined. In terminology proposed by Nicolas Bourbaki 2 in 1938, this is called the dual space of X and 2 Nicolas Bourbaki is the pseudonym of a “secret” society of mathematicians, nearly all French, formed in 1935. It included among its founding members A. Weil, J. Dieudonn´e, and H. Cartan. New members were added over time, and one of its rules was that members were to retire at age

1.4 Linear Functionals

17

is denoted X ∗ . Our immediate goal is to understand the dual space of a Hilbert space. The next result is called the Riesz representation theorem; it was discovered independently by Riesz and Fr´echet in 1907. As motivation for the statement, observe that if we fix a vector h0 in a Hilbert space H , then the map Λ (h) ≡ h, h0 is clearly linear on H . The Cauchy–Schwarz inequality shows that Λ is bounded with Λ  ≤ h0 . In fact we have equality, as is easily seen by computing Λ (h0 /h0 ) if h0 = 0. The Riesz representation theorem provides a converse to these observations. Theorem 1.29. Every bounded linear functional Λ on a Hilbert space H is given by inner product with a (unique) fixed vector h0 in H : Λ (h) = h, h0 . Moreover, the norm of the linear functional Λ is h0 . Proof. Suppose Λ is a bounded linear functional on H . If Λ is identically 0, choose h0 = 0. Otherwise, set M = ker Λ ≡ {h ∈ H : Λ (h) = 0}. Since Λ is linear, M is a subspace of H , and since Λ is continuous, M = Λ −1 (0) is closed. Note that M = H since we are assuming Λ = 0. Pick a nonzero vector z ∈ M ⊥ . By scaling if necessary we may assume Λ (z) = 1. Consider, for arbitrary h ∈ H , the vector Λ (h)z − h and observe that if we apply Λ to this vector we get 0, i.e., it lies in M. Since z was chosen to lie in M ⊥ , this says

Λ (h)z − h ⊥ z so that for every h ∈ H ,

Λ (h)z − h, z = 0.

Rearranging this last line we see that Λ (h) = h, z/z2 , which gives the existence statement with h0 = z/z2 . Uniqueness is immediate, and since we have already observed that Λ  = h0 , we are done. What does the proof of this result tell you about the relationship between any two vectors in (ker Λ )⊥ when Λ is a bounded linear functional on H ? Theorem 1.29 says that a Hilbert space is self-dual, i.e., that H ∗ = H in the sense that the map sending h0 in H to the bounded linear functional ·, h0 is an isometry of H onto its dual space (“isometry” referring to the fact that the norm of the linear functional induced by h0 is h0 ) . Notice that we’re not asserting linearity for the identification of h0 with the linear function ·, h0 ; why not? In their 1907 works, Riesz and Fr´echet dealt specifically with the Hilbert space L2 [a, b]. Shortly thereafter, Riesz considered the natural generalization of his work when he investigated the possibility of describing all bounded linear functionals on L p [a, b] for 1 ≤ p < ∞, launching the study of L p spaces as normed linear spaces. In 1909, Riesz identified the set of all bounded linear functionals on L p [a, b], 1 ≤ p < ∞ 50. The society’s original purpose was to create an analysis text, but this quickly expanded into a ´ ements de math´ematique, now totaling more than project of much bigger scope. A multivolume El´ 7000 pages and treating many core topics in modern mathematics, has been produced.

18

1 Hilbert Space Preliminaries

with Lq [a, b], where 1/p + 1/q = 1 (when p = 1 we set q = ∞). The analogous statement for  p came a few years earlier, in work of E. Landau; the reader is asked to provide a proof in this case in Exercise 1.17. If we leave the realm of Banach spaces, however, a discussion of bounded linear functionals may become moot. For example, M.M. Day showed in 1940 that there are no continuous linear functionals on L p [0, 1] for 0 < p < 1 except the trivial functional (which is identically zero). The spaces L p [0, 1] for 0 < p < 1 are discussed in Exercise 1.30; they are not Banach spaces. Let us return to our example of the Bergman space La2 (D). Observe that Corollary 1.19 says that evaluation at any point w ∈ D is a bounded linear functional on the Hilbert space La2 (D). By Theorem 1.29, evaluation at w must thus be given by inner product with some fixed vector in La2 (D), that is, for each w ∈ D there is a function in La2 (D), which we will denote Kw (z), satisfying f (w) = f , Kw for all f ∈ La2 (D). Can we identify Kw ? This has a nice answer, which is outlined in Exercise 1.25. Next we’ll interpret the projection theorem when H = L2 (D, dA/π ) and M = 2 La (D), the Bergman space. Can we find an explicit formula for the orthogonal projection P : L2 (D, dA/π ) → La2 (D)? A simple lemma will be useful here. Lemma 1.30. Let P : H → M be the orthogonal projection of a Hilbert space H onto a closed subspace M of H . We have f , Pg = P f , g for all vectors f and g in H . Proof. Let f and g be in H and write, using the projection theorem, f = m1 + n1 , g = m2 + n2 , where m1 , m2 ∈ M and n1 , n2 ∈ M ⊥ . We have f , Pg = m1 + n1 , m2 = m1 , m2

while P f , g = m1 , m2 + n2 = m1 , m2 . Returning to our question, if f ∈ L2 (D, dA/π ), then for any w ∈ D, P f (w) = P f , Kw = f , PKw = f , Kw = La2 (D)

 D

f (z)Kw (z)

dA , π

that gives the linear functional of evaluation at where Kw is the vector in w, and we have used the lemma for the second equality. Since by Exercise 1.25 Kw (z) = (1 − wz)−2 , this gives an integral formula for computing the projection P f . The Bergman space furnishes an example of what are called functional Banach spaces. Here is the definition: A Banach space X consisting of scalar-valued functions on a set S is a functional Banach space if point evaluation es ( f ) ≡ f (s) at each point s of S is a bounded linear functional on X, and if no evaluation functional es is identically 0. Other examples of functional Banach spaces, besides La2 (D), include C[0, 1] in the supremum norm and  p for 1 ≤ p ≤ ∞. A non-example is L p ([0, 1], dx), 1 ≤ p ≤ ∞; here the vectors are equivalence classes of functions, and evaluation at a point of [0, 1] doesn’t even make sense.

1.5 Orthonormal Bases

19

1.5 Orthonormal Bases Definition 1.31. An orthonormal set in a Hilbert space H is a set E with the properties: (1) (2)

for every e ∈ E , e = 1, and for distinct vectors e and f in E , e, f = 0.

For an easy example of an orthonormal set in the Hilbert space 2 , take the set E of vectors e j , j ≥ 1 where e j has a 1 in the jth coordinate and zeros elsewhere. As a second example, consider the Hilbert space L2 [0, 2π ], with respect to normalized Lebesgue measure dt/(2π ). The collection of functions eint for any integer n form an orthonormal set in this Hilbert space. We often will write L2 (T ) for L2 ([0, 2π ], dt/(2π )), where T denotes the unit circle and we are identifying a function on [0, 2π ] with a function on T by f (t) = f (eit ). Definition 1.32. An orthonormal basis for a Hilbert space H is a maximal orthonormal set; that is, an orthonormal set that is not properly contained in any orthonormal set. It is easy to see that in the 2 example above, the set {e j : j ≥ 1} is an orthonormal basis. Harder, but still true, is that {eint : n ∈ Z}, where Z is the set of all integers and eint = cos(nt) + i sin(nt), is an orthonormal basis for L2 (T ). This result is a consequence of Fej´er’s theorem; for a proof the reader is referred to [48]. Every Hilbert space has an orthonormal basis (see Exercise 3.1 in Chapter 3). The proof of this statement uses Zorn’s lemma, which will be discussed in Section 3.1. The Hilbert spaces of principal interest to us will either have a finite or countably infinite orthonormal basis. A Hilbert space is also a vector space, and as such it has a linear (or Hamel) basis. We digress here briefly to recall some facts and terminology from linear algebra. Given a nonempty subset S in a vector space V , by a linear combination of vectors in S we mean a finite sum of the form n

∑ α jv j

j=1

where the vectors v j are in S and the coefficients α j are scalars. A set S spans V if every vector in V is a (necessarily finite) linear combination of vectors in S. A set S of vectors is said to be linearly independent if the only linear combination of vectors in S that is equal to the zero vector is the one whose scalar coefficients are all zero. A linear, or Hamel, basis for a vector space V is a subset of V that is both linearly independent and spans V . It is easy to see that a Hamel basis can be equivalently defined as a maximal linearly independent subset of V ; that is, a linearly independent set that is not properly contained in any linearly independent set. Every vector space has a Hamel basis, and for a given vector space, any two Hamel bases can be put in one-to-one correspondence; proofs of these can be provided by the reader, or found, for example, in [14]. Notice that the concept of a

20

1 Hilbert Space Preliminaries

Hamel basis depends only on the linear structure of V , and not the topological structure that comes when the vector space is endowed with a norm or inner product. It is for this reason that in a Hilbert space the concept of an orthonormal basis proves to be more central than that of a Hamel basis, so much so that in the context of Hilbert spaces the term “basis” will always mean “orthonormal basis,” and “dimension” will always refer to the (common) cardinality of any orthonormal basis. In particular, a Hilbert space is said to be finite-dimensional if it has a finite orthonormal basis, and infinite-dimensional otherwise. This convention will not lead to any confusion because of the following two facts: A finite orthonormal set in a Hilbert space H that is not properly contained in any orthonormal set is in fact a Hamel basis for H , and no Hilbert space with a finite Hamel basis can contain an infinite orthonormal set. See Exercise 1.21 for a further exploration of these and related ideas. Given a linearly independent sequence { fn }∞ 1 in a Hilbert space H , there always exists an orthonormal sequence {en }∞ such that 1 span{ f1 , f2 , . . . , fk } = span{e1 , e2 , . . . , ek } for each positive integer k, where “span” denotes the set of linear combinations of the indicated set. An inductive process for constructing the vectors e j , called Gram– Schmidt orthonormalization, is outlined in Exercise 1.19. The last topic of this section is motivated by the question: When is an orthonormal set in a Hilbert space an orthonormal basis? When {ek } is a finite or countably infinite orthonormal set in H , then for every vector h ∈ H we have

∑ | h, ek |2 ≤ h2 ; this is known as Bessel’s inequality. It follows from the observation that the closest vector to h in the linear span of the orthonormal set {e1 , e2 , . . . , en } is ∑n1 h, ek ek (see Exercise 1.21), and the Pythagorean identity of Proposition 1.22. The identity in (e) of the next result is called Parseval’s identity; it is the equality case of Bessel’s inequality. Theorem 1.33. If {en }∞ 1 is an orthonormal sequence in a Hilbert space H , then the following conditions are equivalent: (a) (b) (c) (d) (e) (f)

{en }∞ 1 is an orthonormal basis. If h ∈ H and h ⊥ en for all n, then h = 0. For every h ∈ H , h = ∑∞ 1 h, en en ; equality here means the convergence in the norm of H of the partial sums to h. For every h ∈ H , there exist complex numbers an so that h = ∑∞ 1 an en . 2 2 For every h ∈ H , ∑∞ 1 | h, en | = h . For all h and g in H , ∑∞ 1 h, en en , g = h, g .

Proof. The equivalence of (a) and (b) follows almost immediately from the definition, since if 0 = h and h ⊥ en for all n, then {en }∞ 1 ∪ {h/h} is an orthonormal set.

1.5 Orthonormal Bases

21

Now assume (b) and suppose h ∈ H and let cn = h, en . By Bessel’s inequalk 2 ity we have ∑∞ 1 |cn | < ∞, so that the partial sums sk = ∑n=1 cn en form a Cauchy sequence in H with sk − sm 2 =

k



|cn |2

n=m+1

whenever k > m. By completeness these partial sums must converge in H to some vector s. We claim s = h, which will show (c). For each fixed n, s, en = lim sk , en = lim sk , en = h, en , k→∞

k→∞

where we have used the continuity of the inner product, a consequence of the Cauchy–Schwarz inequality. Thus s − h, en = 0 for all n, and by (b), s = h. The reverse implication, (c) =⇒ (b), is easy, since if (c) holds, and h ⊥ en for all n, then h = ∑∞ 1 h, en en implies that h = 0. Clearly (c) implies (d) and the reverse implication follows from setting fk = ∑kj=1 a j e j and noting as above that h, en = lim fk , en = lim fk , en = an . k→∞

k→∞

Next we show that (c) implies (e). Continuity of the norm shows that if h = k 2 ∑∞ 1 h, en en , then sk  → h where sk is the partial sum ∑n=1 h, en en and sk  = k 2 by the Pythagorean formula. Thus (e) holds. | h, e

| ∑1 n Clearly (f) implies (e) and the reverse implication can be obtained by using the polarization identity to write h, g in terms of h + g2 , h − g2 , h + ig2 and h − ig2 , expanding each of these norms using (e), and computing. 2 Finally, if (e) holds, and h ⊥ en for all n, then h2 = ∑∞ 1 | h, en | = 0, giving (b). When {en }∞ 1 is an orthonormal basis for a Hilbert space H , and h ∈ H , the scalars h, en are called the Fourier coefficients of h with respect to {en }∞ 1 . In this case, the sum in (c) of the above theorem is referred to as the Fourier series of h, relative to the specified orthonormal basis. No countably infinite orthonormal basis can ever be a Hamel basis. Indeed, using Gram–Schmidt orthonormalization we can show something stronger. Suppose that { f1 , f2 , . . .} is a linearly independent sequence in a Hilbert space H . We claim that there is a vector in H which is not a finite linear combination of the f j . The Gram–Schmidt process produces an orthonormal sequence {e1 , e2 , . . .} with span{e1 , e2 , . . . , ek } = span{ f1 , f2 , . . . , fk } for all positive integers k. Write down any sum ∑∞ 1 c j e j where infinitely many of the coefficients c j are nonzero and ∞ 2 < ∞. This sum converges in H to some vector g, since its partial sums |c | ∑1 j form a Cauchy sequence in H . Since g, en = cn , and this is nonzero for infinitely many n, g is not in the span of {e1 , e2 , . . . , ek } for any k, and hence neither is it in the span of { f1 , f2 , . . . , fk } for any k. √ Example 1.34. It is easy to see that if en = n + 1 zn , then {en }∞ n=0 is an orthonormal sequence in the Bergman space La2 (D). Is it an orthonormal basis? It’s

22

1 Hilbert Space Preliminaries

tempting to think we can make short work of answering this question. By part (d) of Theorem 1.33, we need only show that √ every nfunction f in the Bergman ∞ a e = a space can be written as f = ∑∞ ∑0 n n + 1z , and since f is analytic in 0 n n n the disk it √ has a power series expansion f = ∑∞ 0 bn z . But the two equality signs, in ∞ ∞ n n f = ∑0 an n + 1 z and in f = ∑0 bn z , refer to two different kinds of convergence. In the first, we want convergence of the partial sums in the L2 (D, dA/π ) norm, while the second gives us pointwise convergence in D, or, better, uniform convergence on n compact subsets of D of the partial sums of ∑∞ 0 bn z to f . Since this latter type of 2 convergence does not imply L convergence, something more must be done. 2 To that end, we will show that {en }∞ n=0 is an orthonormal basis for La (D) by 2 showing that if f ∈ La (D) and f ⊥ en for all n = 0, 1, 2 . . ., then f = 0. The assumption that f ⊥ en is simply that  D

f (z)zn

dA = 0. π

k We can write f in terms of its power series in D, f (z) = ∑∞ 0 bk z , where the partial sums of this series converge uniformly on compact subsets of D. Fix t < 1 and use this uniform convergence to write     ∞

f (z)zn dA = tD

tD ∞

=



zn dA

k=0

∑ bk

k=0

∑ bk zk 

zk zn dA tD

bn 2n+2 t n+1

since the integral in the penultimate line is 0 unless k = n, in which case it is π t 2n+2 /(n + 1). Now f (z)zn is in L2 (D, dA/π ) ⊆ L1 (D, dA/π ), so we can let t ↑ 1 and use the dominated convergence theorem to see that for any nonnegative integer n,  dA bn . f (z)zn = π n+1 D Thus if this integral is 0 for all n, we see that bn = 0 for all n and hence f = 0 as desired. Specializing the result of (e) of Theorem 1.33 to H = L2 [a, b] we obtain the Riesz–Fischer theorem, named for simultaneous and independent work of Riesz and Ernst Fischer in 1907. More precisely, Fischer showed that L2 [a, b] is complete (see also Section A.3 in the Appendix), while Riesz showed that given a orthonormal basis {en } of L2 [a, b], the map which sends f ∈ L2 [a, b] to the square-summable  sequence {an }, defined by an = ab f en dx, is an isometric linear bijection onto 2 . These two results are individually or collectively referred to as the “Riesz–Fischer theorem”; they are equivalent in the sense that each can be recaptured from the other. Theorem 1.29, the Riesz representation theorem (or as it should be more accurately

1.6 Exercises

23

called, the Fr´echet–Riesz Representation theorem) enters into the mix as well, in the following argument due to Riesz. Suppose we have an orthonormal basis {ei } (known to Riesz, based on a talk given by Erhard Schmidt it 1905, to be countable) for L2 [a, b] and a sequence {ci } of complex numbers with ∑ |ci |2 < ∞. Define Λ on L2 [a, b] by Λ ( f ) = ∑ ci f , ei , 

where f , ei = ab f ei dx. The mapping Λ is linear, and bounded by the Cauchy– Schwarz inequality and Bessel’s inequality. By Theorem 1.29, there exists g ∈ L2 [a, b] such that  b a

f gdx = Λ ( f ) = ∑ ci f , ei

for all f ∈ L2 [a, b]. Setting f = ei gives  b a

b a

ei gdx = ci , and setting f = g gives

|g|2 dx = ∑ | g, ei |2 . 

Now imagine starting with g ∈ L2 [a, b] and defining ci = ab gei dx. Defining Λ as above, we get the stated isometric linear bijection between L2 [a, b] and 2 . Riesz also provided three proofs for the completeness of L2 [a, b]. Exercise 1.27 outlines one of these, which relies on Theorem 1.29 as well.

1.6 Exercises 1.1. Prove the reverse triangle inequality: For vectors x, y in any normed linear space, x + y ≥ |x − y| . 1.2. Show that C[0, 1] is a Banach space in the supremum norm. Hint: If { fn } is a Cauchy sequence in C[0, 1], then for each fixed x ∈ [0, 1], { fn (x)} is a Cauchy sequence in C, which is complete. 1.3. Let C1 [0, 1] be the space of continuous, complex-valued functions on [0, 1] with continuous first derivative. Show that in the supremum norm  · ∞ , C1 [0, 1] is not a Banach space, but that in the norm defined by  f  =  f ∞ +  f  ∞ it does become a Banach space. 1.4. Show that the space 1 of Example 1.5 is complete. 1.5. Show that a metric space is complete if every Cauchy sequence has a convergent subsequence. 1.6. Assume that you know Minkowski’s inequality 

n

∑ |a j + b j |

j=1



1/2 2



n

∑ |a j |

j=1



1/2 2

+

n

1/2

∑ |b j |

j=1

2

24

1 Hilbert Space Preliminaries

for Cn in the Euclidean norm. (a) Show that for {an } and {bn } in a weighted sequence space 2β , 





1/2

∑ |a j + b j | β ( j) 2

2

∑ |a j | β ( j)



j=0



1/2



2

2

j=0

+



1/2

∑ |b j | β ( j) 2

2

.

j=0

(b) Verify directly (without appealing to the Riesz–Fischer theorem on the completeness of L2 (X, µ ) in general) that 2β is complete. 1.7. Let x and y be any two vectors in an inner product space and set λ = y, y . Show that λ λ x, x − | x, y |2 = λ x − x, y y, λ x − x, y y . Use this to derive the Cauchy–Schwarz inequality and to determine when equality holds in the Cauchy–Schwarz inequality. 1.8.(a) Show that for a normed linear space X, the map x → x of X into [0, ∞) is continuous. Is it uniformly continuous? (b) Show that the mappings X × X → X given by (x, y) → x + y, and C × X → X given by (α , x) → α x are continuous. The topologies on X × X and C × X are the product topologies. (c) Suppose that X is an inner product space. Show that the maps x → x, y and x → y, x are continuous on X for each fixed y in X. Are they uniformly continuous? ∞ 2 n 1.9.(a) Show that if ∑∞ 0 |an | < ∞, then the power series ∑0 an z has radius of convergence at least one, and hence is an analytic function in the unit disk D. Hint: Recall that the radius of convergence R is determined by

1 = lim sup |an |1/n . R n→∞ (b) Show that if lim β (n)1/n ≥ 1

n→∞

and



∑ |an |2 β (n)2 < ∞ 0

then the power series



∑ an zn 0

has radius of convergence at least equal to 1.

1.6 Exercises

25

1.10. Suppose that X and Y are normed linear spaces and that T : X → Y is a linear map (meaning that T (α x1 + β x2 ) = α T (x1 ) + β T (x2 ) for all vectors x1 , x2 in X and all scalars α and β ). Suppose that T maps X onto Y and is isometric (meaning T x = x for all x ∈ X). (a) Show that T is one-to-one. (b) Show that if X is a Banach space, so is Y . (c) Show that if X is a Hilbert space, then so is Y if we define y1 , y2 Y = x1 , x2 X where x1 and x2 are the unique points in X satisfying T x1 = y1 and T x2 = y2 . (d) Explain how this shows that the weighted Hardy spaces H 2 (β ) of Example 1.7 are Hilbert spaces. 1.11. For 1 ≤ p < ∞, define Lap (D) to be the set of all analytic functions f on the unit disk D for which  dA | f (z)| p 0. The idea is to consider something like χE to show that Mϕ  > α , but we can’t quite do this since χE will not be in L2 (X, µ ) if µ (E) is infinite. Instead, we use the σ -finiteness hypothesis on µ to find a subset E  of E with 0 < µ (E  ) < ∞. Then f = χE  will be in L2 (X, µ ) and Mϕ f 2 =

 E

|ϕ |2 d µ > α 2 µ (E  ) = α 2

 X

| f |2 d µ

so that Mϕ  > α , where α is any chosen value less that ϕ ∞ . Thus Mϕ  = ϕ ∞ . In the next example, which combines the previous two, we encounter the composition of two linear operators. In general, if A ∈ B(X,Y ) and B ∈ B(Y,V ), then by the product BA we mean the mapping defined by BA(x) = B(A(x)) for x ∈ X. It is easy to see that BA ∈ B(X,V ) and BA ≤ BA. Example 2.6. Consider the Bergman space La2 (D) and let ϕ ∈ L∞ (D, dA/π ); note that ϕ is not assumed to be analytic. Define the Toeplitz operator with symbol ϕ on La2 (D) by Tϕ ( f ) = P(ϕ f ), where P is the projection of L2 (D, dA/π ) onto La2 (D).

2.1 Bounded Linear Operators

33

Clearly Tϕ is linear, and since P : L2 (D, dA/π ) → La2 (D) is bounded with norm 1, and Mϕ : L2 (D, dA/π ) → L2 (D, dA/π ) is bounded of norm ϕ ∞ , Tϕ is bounded with norm at most ϕ ∞ . When ϕ is analytic in L∞ (D, dA/π ), the projection factor in the definition of Tϕ serves no purpose, and in this case Tϕ is just the restriction of the multiplication operator Mϕ to the closed subspace La2 (D) of L2 (D). Example 2.7. The next pair of operators are simple but important ones. They act from 2 to itself. The first, called the forward shift, is defined by S(x1 , x2 , . . .) = (0, x1 , x2 , . . .). It is easy to see that it is a bounded linear operator of norm one; in fact it is an isometry, meaning Sx = x for every x = (x1 , x2 , . . .) ∈ 2 . The backward shift is the operator from 2 to 2 which takes (x1 , x2 , x3 , . . .) to (x2 , x3 , . . .). It has norm 1, but is not an isometry (why?). Example 2.8. Suppose that H is a Hilbert space with orthonormal basis {en }∞ 1. Choose any bounded sequence of complex numbers {αn }∞ 1 and set Aen = αn en . Extend A by linearity to any finite linear combination of the en , and extend A to all of H by continuity, noting that linearity is preserved. Explicitly this means the following. Given h ∈ H , we know h = ∑∞ 1 h, en en , where the sum converges in H . Since {αn } is a bounded sequence, the partial sums of ∑∞ 1 h, en αn en form a Cauchy sequence in H and thus converge in H ; call the sum Ah. Since ∞



Ah2 = ∑ | h, en |2 |αn |2 ≤ (sup |αn |2 ) ∑ | h, en |2 = (sup |αn |)2 h2 , n

1

1

n

we see that A is bounded with A ≤ supn |αn |. Consideration of Aen shows that, indeed, we have equality here. Such an operator A is called a diagonal operator, with diagonal sequence {αn }. The terminology comes from defining, in analogy with the finite-dimensional case, the matrix MA of A (with respect to the basis {en }) to be the (infinite) matrix with i jth entry Ae j , ei . This is a diagonal matrix when A is a diagonal operator. In spite of its usefulness in finite-dimensional settings, we will not find it particularly helpful to work with the matrix of a general bounded linear operator on a Hilbert space H , in part because it is not easy to tell if a linear operator A is bounded by looking at its matrix MA . Another way to say this is that while every bounded linear operator corresponds to a matrix, the converse is not true. Example 2.9. We describe a class of operators called integral operators. Start with a σ -finite measure space (X, M, µ ) and a measurable function k : X × X → C with k ∈ L2 (X × X, µ × µ ). Define K : L2 (X, µ ) → L2 (X, µ ) by K f = g where 

g(x) = X

k(x, y) f (y)d µ (y)

for x in X. We call k the kernel of the integral operator K. Since k ∈ L2 (X × X), for almost every x ∈ X, the function y → k(x, y) is in L2 (X, µ ), and thus the function

34

2 Operator Theory Basics

y → k(x, y) f (y) is integrable if f ∈ L2 (X, µ ). We show that K maps into L2 (X, µ ) and is bounded. For f ∈ L2 (X, µ ) we have  

2   K( f ) = |g(x)| d µ (x) =  k(x, y) f (y)d µ (y) d µ (x) X X X 2   ≤ |k(x, y)|| f (y)|d µ (y) d µ (x) X X      |k(x, y)|2 d µ (y) | f (y)|2 d µ (y) d µ (x) ≤ 

2

2

X

X

=  f 2

 

X X 2 2

X

|k(x, y)|2 d µ (y)d µ (x)

=  f  k ,

where we have used the Cauchy–Schwarz inequality midway through the calculation. This computation shows that K is bounded from L2 (X, µ ) into L2 (X, µ ) with norm at most k. Some of the measure-theoretic technicalities of this example can be bypassed by taking, for example, X = [0, 1] and requiring k(x, y) to be continuous on [0, 1] × [0, 1]. A particular integral operator of interest is the Volterra operator; it comes from the choice X = [0, 1], with Lebesgue measure, and k(x, y) equal to the characteristic function of the lower triangle {(x, y) : y ≤ x} in the unit square [0, 1] × [0, 1]. This gives  x

f (y)dy,

K f (x) = 0

so the Volterra operator is sometimes called the √ “operator of indefinite integration.” By the above remarks, its norm is at most 1/ 2; computing the norm exactly is not so easy; see [17], Problem 188. A bounded linear operator T on a Banach space X is said to attain its norm if there is a nonzero vector x in X with T x = T x. See Exercise 2.8 for an exploration of this issue in the Hilbert space setting.

2.2 Adjoints of Hilbert Space Operators Now that we have some examples of bounded linear operators in mind, let us turn to the notion of the adjoint of a Hilbert space operator. Later we will define adjoints of operators on Banach spaces, and compare it to the definition we give now in the Hilbert space setting. As motivation for our work in this section, recall that given an n × n matrix A = (ai j ) with complex entries, its conjugate transpose A∗ is the n × n matrix whose i jth entry is a ji . Associate to the matrix A the linear operator TA on Cn given by TA (v) = Av where v ∈ Cn is written as a column vector. For any vectors v and w in Cn , we have

2.2 Adjoints of Hilbert Space Operators

TA v, w = v, TA∗ w .

35

(2.1)

The operator TA∗ is called the adjoint of the operator TA , and the analogue of the property in Equation (2.1) will lead us to the idea of the adjoint of any bounded linear operator between Hilbert spaces. To make this precise, we begin with the definition of a sesquilinear form. Definition 2.10. If H and K are both Hilbert spaces, a sesquilinear form u : H × K → C is a mapping satisfying (1) (2)

u(α h + β g, k) = α u(h, k) + β u(g, k), and u(h, α k + β f ) = α u(h, k) + β u(h, f )

for all h, g ∈ H , all k, f ∈ K and all scalars α and β . A sesquilinear form u is bounded if there is a finite constant M such that |u(h, k)| ≤ Mhk for all h ∈ H and k ∈ K . Setting H = K and letting u(h, k) be the inner product h, k gives an example of a sesquilinear form that is bounded (by the Cauchy– Schwarz inequality). More generally, if A ∈ B(H , K ) and B ∈ B(K , H ), then both u(h, k) ≡ Ah, k and u(h, k) ≡ h, Bk (where the inner products are in K and H , respectively) define sesquilinear forms that are bounded, since, for example, |u(h, k)| = | Ah, k | ≤ Ahk ≤ A h k. The next result describes all bounded sesquilinear forms. Theorem 2.11. Let H and K be Hilbert spaces and suppose that u : H × K → C is a bounded sesquilinear form. There exists a unique A ∈ B(H , K ) such that u(h, k) = Ah, k K for all h ∈ H and k ∈ K . Proof. For fixed h ∈ H we define a mapping Λh : K → C by

Λh (k) = u(h, k). One can easily check that Λh is linear. Moreover, since u is bounded by hypothesis, |Λh (k)| = |u(h, k)| = |u(h, k)| ≤ Mhk for some M independent of h and k. Thus Λh is a bounded linear functional on K . By the Riesz representation theorem this functional must therefore be given by inner product with a unique vector f in K :

Λh (k) = k, f K for all k ∈ K , and moreover Λh  =  f . Since we have already observed that Λh  ≤ Mh, we must have  f  ≤ Mh. This process defines a map A from H

36

2 Operator Theory Basics

to K taking h to f . This mapping is linear: If h1 , h2 are in H and α ∈ C with Ah1 = f1 and Ah2 = f2 so that u(h1 , k) = k, f1 K and u(h2 , k) = k, f2 K for all k ∈ K , then

Λα h1 +h2 (k) = u(α h1 + h2 , k) = α u(h1 , k) + u(h2 , k) = α k, f1 K + k, f2 K = k, α f1 + f2 K , so that A maps α h1 + h2 to α f1 + f2 . We have already seen that A is bounded with Ah =  f  ≤ Mh. This shows that there is a bounded linear operator A with u(h, k) = k, Ah K or equivalently u(h, k) = Ah, k K . Moreover, A is unique, since if f1 , k K = f2 , k K for all k ∈ K , we must have f1 − f2 , k K = 0 for all k, and therefore f1 = f2 . As a consequence of the last result, suppose we start with an operator A in B(H , K ) and define u : K × H → C by u(k, h) ≡ k, Ah K . This is a bounded sesquilinear form. Applying Theorem 2.11, we can find the unique operator, call it A∗ , in B(K , H ) satisfying u(k, h) = A∗ k, h H for all k ∈ K and h ∈ H . Taking conjugates, we have the following important conclusion. Theorem 2.12. Given Hilbert spaces H and K and A ∈ B(H , K ), there is a unique A∗ ∈ B(K , H ) so that Ah, k K = h, A∗ k H for all h ∈ H and k ∈ K . The operator A∗ in the last result is called the (Hilbert space) adjoint of A. In the case that H = K and A∗ = A we say that A is self-adjoint or Hermitian. Looking back at the statement of Lemma 1.30 in Chapter 1, we see that orthogonal projections (onto closed subspaces) are self-adjoint operators. Some more examples are in order. For the forward shift S on 2 as in Example 2.7, it is easy to check that S∗ is the backward shift; to see this it suffices to compute

2.2 Adjoints of Hilbert Space Operators

37

Sx, y and x, By and see that they agree, where x and y are in 2 and B denotes the backward shift. For a multiplication operator Mϕ , defined on L2 (X, µ ) for some σ -finite measure space (X, µ ) and ϕ ∈ L∞ (X, µ ), we have Mϕ f , g = ϕ f , g =

 X

ϕ f gd µ = f , ϕ g = f , Mϕ g

for any f , g ∈ L2 (X, µ ). Thus Mϕ∗ = Mϕ , and a multiplication operator is self-adjoint if and only if its symbol ϕ is real-valued almost everywhere. A similar computation, using the self-adjointness of the Bergman projection from Lemma 1.30, shows that a Toeplitz operator Tϕ on La2 (D) has adjoint Tϕ . See Exercise 2.5. To determine the adjoint of the integral operator with kernel k acting on L2 (X, µ ) we seek the operator K ∗ satisfying K f , g = f , K ∗ g for all f and g in L2 (X, µ ). Writing the inner product as an integral, and using Fubini’s theorem to interchange the order of integration, we see that K f , g =



K f (x)g(x)d µ (x)

X  

 k(x, y)g(x)d µ (x) f (y)d µ (y) = X X    k(x, y)g(x)d µ (x) f (y)d µ (y), = X

which is equal to

 X

X

f (y)K ∗ g(y)d µ (y) = f , K ∗ g

if we define K ∗ to be the integral operator with kernel k∗ (x, y) ≡ k(y, x). We will be mainly interested in adjoints for bounded linear operators from a Hilbert space H to itself; recall we will write B(H ) for B(H , H ) in this case. For simplicity, the next result is stated in this setting, rather than the more general one of bounded linear operators from H to K . Proposition 2.13. For A and B in B(H ) we have (a) (b) (c) (d)

A∗∗ = A where A∗∗ = (A∗ )∗ . (A + B)∗ = A∗ + B∗ . (α A)∗ = α A∗ for α ∈ C. (AB)∗ = B∗ A∗ .

Proof. For (a) we first note that by the definition of the adjoint we have A∗ x, y = x, A∗∗ y for all x and y in H . Since also Ay, x = y, A∗ x , taking conjugates we see that x, Ay = A∗ x, y . Thus x, A∗∗ y = x, Ay for all x and y, or equivalently x, A∗∗ y − Ay = 0 for all x and y. This forces A∗∗ y = Ay for all y ∈ H , giving (a).

38

2 Operator Theory Basics

Part (b) follows from the definition of the adjoint and a straightforward computation showing that (A + B)x, y = x, (A∗ + B∗ )y for all x and y in H . Parts (c) and (d) are done similarly, and the details are left to the reader. The next result will be fundamentally important to us. Proposition 2.14. If A ∈ B(H ), then A = A∗  and A∗ A = A2 . Proof. Take any vector h ∈ H with h = 1. We have Ah2 = Ah, Ah = h, A∗ Ah ≤ A∗ Ah h ≤ A∗ A ≤ A∗  · A

(2.2)

so that A ≤ A∗ . Applying this together with (a) of the previous proposition we also have A∗  ≤ A∗∗  = A. Thus A = A∗ . Using this, and taking the supremum over all unit vectors h in (2.2), we see that A2 = sup{Ah2 : h = 1} ≤ A∗ A ≤ A∗  · A = A2 and thus equality must hold throughout, yielding A2 = A∗ A as desired.



Let us recap and extend the structure we have on B(H ) when H is any Hilbert space. First of all, using pointwise-defined vector operations and the “operator norm,” it is a Banach space; this is Proposition 2.3. We define a multiplication on B(H ) which makes it into a complex algebra; that is, a vector space over C with a multiplication satisfying A(BC) = (AB)C, (A + B)C = AC + BC, A(B + C) = AB + AC, and α (AB) = (α A)B = A(α B) for all A, B,C ∈ B(H ) and scalar α . This multiplication AB is just the composition of the linear maps A and B. Note multiplication is not in general commutative. As we have noted above, AB ≤ AB, and we will see later this makes B(H ) into a Banach algebra, as will be formally defined in Chapter 5. We also have an involution ∗ of B(H ); this is a map A → A∗ of B(H ) into itself satisfying (A∗ )∗ = A, (AB)∗ = B∗ A∗ , (α A + B)∗ = α A∗ + B∗ for all A, B in B(H ) and scalars α . Of course, ∗ is just our adjoint operation. It is connected to the norm by A∗ A = A2 , the result of Proposition 2.14. As we will see in Chapter 5, all of these properties together say that B(H ) is a C∗ -algebra. The relationship A∗ A = A2 is called the C∗ -identity. Because there is a multiplicative identity (the identity operator I in B(H )), we will say that B(H ) is a unital C∗ -algebra. Later we will make a study of C∗ -algebras in general, and B(H ) will be one of our primary examples. Since B(H ) is noncommutative (except in the trivial situation that the dimension of H is 1), we will find it convenient to single out for special scrutiny the operators in B(H ) which do commute with their adjoints. Definition 2.15. An operator A in B(H ) is normal if AA∗ = A∗ A, and self-adjoint if A = A∗ . Any multiplication operator Mϕ on L2 (X, µ ) is normal, since Mϕ∗ = Mϕ . In Chapter 6, we will see multiplication operators are, in a natural sense, the canonical examples of normal operators.

2.2 Adjoints of Hilbert Space Operators

39

As noted above, we will be primarily interested in adjoints of operators in B(H ), as opposed to B(H , K ). One exception is if H and K are Hilbert spaces and U : H → K is a linear surjection that preserves inner products, meaning Uh1 ,Uh2 = h1 , h2 for all h1 and h2 in H . Such a map is called a Hilbert space isomorphism. Proposition 2.16. If U : H → K is an isomorphism, then U ∗U = IH (the identity on H ) and UU ∗ = IK . Proof. Let h and g be in H . We have U ∗Uh, g = Uh,Ug = h, g

so that U ∗Uh − h, g = 0. This says that for a fixed h, U ∗Uh − h is orthogonal to every vector in H , and hence U ∗U = IH . Now let k be in K . Since U is surjective we may find h with Uh = k. Thus UU ∗ k = (UU ∗ )Uh = Uh = k, which gives the desired statement about UU ∗ . Definition 2.17. An operator A in B(H , K ) is said to be invertible if there exists B in B(K , H ) with AB = IK and BA = IH . We write B = A−1 . We can rephrase the last proposition as “If U is an isomorphism, then U ∗ = U −1 .” An isomorphism between Hilbert spaces is called a unitary operator, although some authors will restrict this terminology to the setting H = K , and some, as is our practice, will use it more generally for an isomorphism of one Hilbert space H onto another, possibly different, Hilbert space K . Definition 2.18. If H and K are Hilbert spaces and if U : H → K is a bijective linear map with Uh1 ,Uh2 K = h1 , h2 H for all h1 and h2 in H , then U is said to be a unitary operator. It is an easy consequence of the polarization identity that a linear and surjective isometry from H to K is unitary; see Exercise 2.9. Let us make a few elementary observations about invertibility of operators. First note that when A ∈ B(H , K ) is invertible with inverse B, A must be one-to-one, since if Ah1 = Ah2 we must have h1 = BAh1 = BAh2 = h2 . The operator A must also be onto K , since for any k ∈ K , A(Bk) = k. We’ll discuss a converse to these statements a bit later. By linearity, A is one-to-one if and only if its kernel ker A ≡ {h ∈ H : Ah = 0} consists of the zero vector only. We can add another property to our list of properties of ∗ : Proposition 2.19. If A is invertible, then so is A∗ , and (A∗ )−1 = (A−1 )∗ .

40

2 Operator Theory Basics

Proof. When A is invertible we have AA−1 = I = A−1 A. Applying the ∗ operation we have (AA−1 )∗ = I ∗ = (A−1 A)∗ ; clearly I ∗ = I so that by property (d) of Proposi tion 2.13 we have (A−1 )∗ A∗ = I = A∗ (A−1 )∗ , which is the desired conclusion. The next result is the converse to Proposition 2.16. Proposition 2.20. If U is in B(H , K ) with U invertible and U −1 = U ∗ , then U is an isomorphism. Proof. We have already observed that U must be surjective, so we only need to check that it preserves inner products: Uh,Ug = h,U ∗Ug = h,U −1Ug = h, g

for all h and g in H .



A few examples are in order. Example 2.21. Let S be the forward shift on 2 , so that S∗ is the backward shift. We have S∗ S = I, but S is not unitary, since its not surjective. This example points out an important distinction with the finite-dimensional situation. For a linear map T from Cn into itself, T is necessarily bijective if it is either one-to-one or surjective. Example 2.22. Consider a multiplication operator Mϕ on L2 (X, µ ) for ϕ in L∞ (X, µ ). When is Mϕ unitary? We want Mϕ Mϕ−1 = Mϕ Mϕ∗ = I. We know that Mϕ∗ = Mϕ , so that Mϕ is unitary if and only if |ϕ |2 f = f for all f in L2 (X, µ ); that is, if and only if |ϕ | = 1 µ -almost everywhere. Example 2.23. Let F map L2 ([0, 2π ], dt/(2π )) into 2 (Z) ≡ {{an }∞ −∞ :





|an |2 < ∞}

n=−∞

by F( f ) = { fˆ(n)}∞ −∞ where fˆ(n) = f , eint =

 2π 0

f (t)e−int

dt . 2π

Linearity of F follows from linearity of the integral. Since {eint } is an orthonormal basis for L2 ([0, 2π ], dt/(2π )), part (f) of Theorem 1.33 guarantees that F preserves ∞ 2 int inner products. Given a sequence {an }∞ −∞ ∈  (Z), define f ≡ ∑−∞ an e . This sum 2 ∞ converges in L ([0, 2π ]) and F( f ) = {an }−∞ . Thus F is surjective, and hence is a unitary map. Suppose A ∈ B(H , K ), and think of A as a mapping from H to K in the purely set-theoretic sense. From this point of view, A is invertible (as a mapping between sets) if and only if A is bijective. Its not hard to show (see Exercise 2.7) that if A is bijective, linearity of A implies that this set-theoretic inverse is a linear

2.3 Adjoints of Banach Space Operators

41

map from K onto H . However, it is not at all clear that this linear set-theoretic inverse should be bounded. Remarkably, it is, as we will see in Section 3.3, and we are left with the extremely useful conclusion that a bounded linear operator in B(H , K ) is invertible (in the sense of Definition 2.17) if and only if it is bijective. We explore this a bit further. Our immediate goal is to show that we can characterize the invertible operators by weakening the requirement that A be onto if we simultaneously strengthen the requirement that A be one-to-one. The next definition describes this strengthening. Definition 2.24. We say that A ∈ B(H , K ) is bounded below if there is a δ > 0 such that Ah ≥ δ h for all h in H . Clearly, if A is bounded below, its kernel is {0} and hence A is one-to-one. However, A being one-to-one does not imply that A is bounded below; a diagonal operator with diagonal sequence {1/n} provides a counterexample. A weakening of the condition “A maps H onto K ” is the requirement that the range of A is dense in K ; i.e., the closure of the range of A should be all of K . Theorem 2.25. If A is a bounded linear operator from a Hilbert space H to a Hilbert space K , then A is invertible if and only if A is bounded below and has dense range. Proof. The “only if” direction is easy: A invertible guarantees that the range of A is equal to K , and moreover for any h ∈ H , h = A−1 Ah ≤ A−1 Ah so that Ah ≥

1 h A−1 

and A is bounded below. The “if” direction is outlined in Exercise 2.12.



When we ask why a particular operator fails to be invertible, it is sometimes more useful to see which of the properties “bounded below” and/or “dense range” it fails to have, rather than looking at the properties “one-to-one” and “onto.”

2.3 Adjoints of Banach Space Operators So far we have defined A∗ when A is a bounded linear operator between Hilbert spaces, and our definition seems closely tied to the inner product structure. We pause briefly in this section to see if we can define the adjoint of a bounded linear operator between Banach spaces. For simplicity, we restrict attention to A ∈ B(X), where X is a Banach space. Let us begin by rephrasing the defining property of the adjoint of a Hilbert space operator. If A ∈ B(H ) where H is a Hilbert space, then A∗ is the unique bounded linear operator on H satisfying

42

2 Operator Theory Basics

Ax, y = x, A∗ y

(2.3)

for all x and y in H . Let the linear functional x → x, y on H (for fixed y) be denoted by Λy . Thus we can rewrite Equation (2.3) as

Λy (Ax) = ΛA∗ y (x). Even more suggestively, let us think of the map that associates the linear functional Λy to the linear functional ΛA∗ y . We have

ΛA∗ y (x) = x, A∗ y = Ax, y = Λy (Ax) = (Λy ◦ A)(x) for x and y in H . This suggests we try defining A∗ for A ∈ B(X), where X is now a Banach space, as the map on the dual space X ∗ that sends Λ ∈ X ∗ to Λ ◦ A: A∗ (Λ ) = Λ ◦ A. It is easy to see that A∗ (Λ ) is in X ∗ and that the map A∗ : X ∗ → X ∗ is linear. If we adopt this as the definition of A∗ when A is a bounded linear operator on the Banach space X, how does it compare with our earlier definition in the case that X is actually a Hilbert space H ? To answer this, let C : H → H ∗ be the surjective conjugate linear isometry sending y to Λy ; “conjugate linear” referring to the fact that C(α y) = α C(y) for scalars α . We claim that CA∗HS = A∗BSC, as schematically illustrated below. Here the subscripts HS and BS indicate we are using the adjoint definition in, respectively, the Hilbert space setting or Banach space setting, so that A∗HS acts on H while A∗BS acts on H ∗ . H

A∗HS

C

? H∗

- H C

A∗BS

? - H∗

This is verified by observing that CA∗HS x is the bounded linear functional on H given as inner product with A∗HS x, that is, the bounded linear functional taking y ∈ H to y, A∗HS x = Ay, x . On the other hand, A∗BSCx is the bounded linear functional on H taking y to [A∗BS (Cx)](y) = Cx(Ay) = Ay, x . The conjugate linearity of C means that while (α AHS )∗ = α A∗HS , (α A)∗BS = α A∗BS . It is pleasant to import the Hilbert space notation ·, · into the Banach space setting. For X a Banach space, denote a generic element of X ∗ by x∗ and write x∗ (x) ≡ x, x∗ . Notice, then, that if A is in B(X) and A∗ is in B(X ∗ ), we have Ax, x∗ = x, A∗ x∗

since both are equal to x∗ (Ax).

2.4 Exercises

43

Example 2.26. Consider the Banach space X = C[0, 1] in the supremum norm. As in Exercise 2.3, let ϕ be a continuous map of [0, 1] into [0, 1] and define the bounded linear operator Cϕ on X by Cϕ ( f ) = f ◦ ϕ . Fix a point p ∈ [0, 1] and let Λ p be the bounded linear functional of evaluation at p: Λ p ( f ) = f (p) for all f ∈ X. We seek to identify Cϕ∗ (Λ p ) as an element of X ∗ . We have Cϕ∗ (Λ p )( f ) = Λ p (Cϕ ( f )) = Λ p ( f ◦ ϕ ) = f (ϕ (p)) = Λϕ (p) ( f ) for every f in X. Thus Cϕ∗ (Λ p ) = Λϕ (p) . In our “inner product” notation the relevant calculation looks like f ,Cϕ∗ (Λ p ) = f ◦ ϕ , Λ p = f (ϕ (p)) = f , Λϕ (p) . The concept of the adjoint operator had its beginnings in the work of Riesz in 1909 for operators on L p [a, b]. Riesz used the terminology “Transponierte” or “transposed operator.” By 1930 the idea had been extended by Banach and Juliusz Schauder to the general setting of a bounded linear operator between Banach spaces, with the terms “op´eration adjointe” and “op´eration conjugu´ee” being introduced.

2.4 Exercises 2.1. Let X and Y be normed linear spaces, and let B(X,Y ) denote the collection of all bounded linear operators from X into Y endowed with the operator norm. Show that B(X,Y ) is a normed linear space, and B(X,Y ) is a Banach space whenever Y is a Banach space. The vector operations in B(X,Y ) are to be defined pointwise: (A + B)(x) = Ax + Bx, and (α A)(x) = α (Ax). 2.2. Suppose M is a dense subspace in a Banach space X (meaning that the closure of M is all of X) and suppose that T : M → Y is linear, where Y is a Banach space, with T mY ≤ KmX for some K < ∞ and all m ∈ M. Show that T extends, in a unique way, to a bounded linear operator from X into Y . 2.3. Let X = [0, 1] or more generally any compact Hausdorff space, and let Y = C(X), the Banach space of continuous, complex-valued functions on X, in the supremum norm. For any continuous function ϕ mapping X into itself, define the composition operator Cϕ on Y by Cϕ ( f ) = f ◦ ϕ . Prove that Cϕ is a bounded linear operator on Y . For which ϕ is Cϕ invertible? 2.4. Compute the norm of the multiplication operator Mz (equivalently the Toeplitz operator Tz ) on La2 (D). 2.5. Show that the Toeplitz operator with symbol ϕ acting on the Bergman space La2 (D) has adjoint Tϕ .

44

2 Operator Theory Basics

2.6. Suppose T is a bounded linear operator on a Hilbert space H and suppose further that the range of T is one-dimensional. Show that there are vectors x and y in H so that T z = z, x y for all z ∈ H . This operator is sometimes written as y ⊗ x. Identify T ∗ in this case. 2.7. Show that if T : X → Y is a bijective linear map, then the set-theoretic inverse T −1 is also linear. 2.8.(a) Suppose that h is a nonzero vector in a Hilbert space H . Show that T ∈ B(H ) attains its norm at h (meaning T h = T h) if and only if T ∗ T h = T 2 h. (b) Extend (a) to the case that T is a bounded linear operator from a Hilbert space H to a Hilbert space K , and then use this result to provide another proof of Proposition 2.16. 2.9. Show that a linear surjective isometry from one Hilbert space H to another Hilbert space K is unitary. 2.10. Let {an }∞ 1 be a bounded sequence of complex numbers. Fix an orthonormal basis {gn } for 2 . The unique linear operator W satisfying W (gn ) = αn gn+1 for all n is called a weighted shift. (a) Find W  and W ∗ . (b) Suppose {an } and {bn } are bounded sequences with |an | = |bn | for all n. Let W and V be the associated weighted shifts. Show that there is a unitary U : 2 → 2 with U −1WU = V . We say that W and V are unitarily equivalent. ∞ 2 2 n 2.11.(a) Show that if ∑∞ n=0 |an | (n!) < ∞, then the power series ∑n=0 an z con∞ n verges for all z ∈ C and hence f (z) ≡ ∑n=0 an z is analytic in C. Define the vector space of entire functions

V ≡ {f =





n=0

n=0

∑ an zn : ∑ |an |2 (n!)2 < ∞}

and put an inner product on V by setting f , g =



∑ an bn (n!)2

n=0

∞ n n when f (z) = ∑∞ n=0 an z and g(z) = ∑n=0 bn z . Show that V is a Hilbert space. 2 (b) Let U : H → V by ∞ an U f = ∑ zn n=0 n! ∞ n 2 2 when f (z) = ∑∞ n=0 an z is in H (so that ∑n=0 |an | < ∞; see Example 1.7). Recalling that the power series coefficients an of f are given by

2.4 Exercises

45

an =

f (n) (0) n!

show that U is a unitary map. (c) Define a linear map D on V by D f = f  , and show that D is a bounded linear operator on V . (d) Let B : H 2 → H 2 be defined by Bf =

f − f (0) . z

Show that B is a bounded linear operator on H 2 and D = UBU −1 , so that D and B are unitarily equivalent. 2.12. Suppose that A ∈ B(H , K ) where A is bounded below and has dense range in K . Show that A is invertible. Hint: Start by showing that if A is bounded below, then the range of A is closed. 2.13. Suppose An ∈ B(Hn ) for n = 1, 2, 3, . . ., where each Hn is a Hilbert space. Assume further that supn An  < ∞. Define A on H ≡ ∑ ⊕Hn by A(h) = A(h1 , h2 , . . .) = (Ah1 , Ah2 , . . .). Show that A ∈ B(H ) and A = supn An . We call A the direct sum of the operators {An } and denote it ∑ ⊕An . 2.14. Suppose that u is a bounded sesquilinear form on H × H for some Hilbert space H . (a) Define the quadratic form uˆ : H → C by u(h) ˆ = u(h, h). Show the polarization identity         1 1 1 1 u(h, g) = uˆ (h + g) − uˆ (h − g) + iuˆ (h + ig) − iuˆ (h − ig) 2 2 2 2 for all h and g in H . (b) Show that if u1 and u2 are bounded sesquilinear forms on H × H with u1 (h, h) = u2 (h, h) for all h in H then u1 and u2 agree on H × H . 2.15. Recall the space C1 [0, 1] defined by C1 ([0, 1]) ≡ { f ∈ C[0, 1] : f  exists and is continuous on [0, 1]}. (a) Recall from Exercise 1.3 of Chapter 1 that C1 [0, 1] is a Banach space if we norm it by  f C1 = max{| f (x)| : 0 ≤ x ≤ 1} + max{| f  (x)| : 0 ≤ x ≤ 1}. Show that for each x ∈ [0, 1], the functional

46

2 Operator Theory Basics

ev1x ( f ) = f  (x) is a bounded linear functional on (C1 [0, 1],  · C1 ). (b) Fix ϕ : [0, 1] → [0, 1] in C1 [0, 1] and define Cϕ on C1 [0, 1] by Cϕ ( f ) = f ◦ ϕ . Show that Cϕ is a bounded linear operator on C1 [0, 1]. What is Cϕ∗ (ev1x )? 2.16. For an operator T in B(H ), where H is a Hilbert space, show that ker T = (ran T ∗ )⊥ , where ker T = {h : T h = 0} and ran T ∗ ≡ range T ∗ = {T ∗ h : h ∈ H }. 2.17. If T is a bounded and self-adjoint operator on a Hilbert space and T 2 = T , show that T is the orthogonal projection onto its range. 2.18. Suppose that PM1 and PM2 are orthogonal projections onto the closed subspaces M1 and M2 of a Hilbert space. (a) Show that PM1 PM2 is an orthogonal projection if and only if PM1 PM2 = PM2 PM1 . (b) If the condition in (a) is satisfied so that PM1 PM2 = PM for some closed subspace M, identify M. 2.19. Show that a diagonal operator on a Hilbert space is an orthogonal projection if and only if its diagonal consists of 0’s and 1’s. 2.20. Fix vectors h1 , h2 , . . . , hn in a Hilbert space H . (a) Define B : Cn → H by n

B(z1 , z2 , . . . , zn ) =

∑ z jh j.

j=1

Calculate B∗ : H → Cn . (b) What is the relationship between the operator B in (a), the n × n matrix A = [ai j ] with ai j = h j , hi , and the operator C : H → H given by n

Ch =

∑ h, h j h j ?

j=1

(c) Show that h1 , h2 , . . . , hn are linearly independent in H if and only if the matrix A is invertible. 2.21. Suppose that T is an operator in B(H ) for some Hilbert space H and suppose that T = T −1 and T = T ∗ . Show that the sets H1 ≡ {h + T h : h ∈ H } and H2 ≡ {h − T h : h ∈ H } are closed subspaces of H with H = H1 ⊕ H2 , and that the restriction of T to H1 is the identity I while the restriction of T to H2 is −I. Conversely, show that if H is the direct sum of two subspaces H1 and H2 with T (h) = h for h ∈ H1 and T (h) = −h for h ∈ H2 , then T = T −1 and T ∗ = T .

2.4 Exercises

47

2.22. If A ∈ B(H ) and h ∈ H , then {h, Ah, A2 h, A3 h, . . .} is called the orbit of h under A. When the orbit of h spans H — that is, when the set of linear combinations of vectors in the orbit of h is dense in H —we call h a cyclic vector for A, and say that A is a cyclic operator. (a) Show that the constant function 1 is a cyclic vector for Mz on La2 (D). (b) Consider the operator from C3 to C3 with matrix ⎡ ⎤ 200 A = ⎣ 0 2 0 ⎦. 001 Show that this operator has no cyclic vector. Cyclic operators will be studied further in Chapter 6. 2.23. Thus far, the only topology we have considered on B(H ) for H a Hilbert space is the topology that comes from the operator norm; this is called the norm topology or the uniform operator topology. However, there are other useful topologies on B(H ), and in this problem we introduce two of them by discussing sequential convergence in two new senses. Definition. Given {Tn } in B(H ), we say Tn → T ∈ B(H ) in the strong operator topology if Tn h → T h for each h ∈ H . This is abbreviated Tn → T (SOT). We say Tn converges to T in the weak operator topology, denoted Tn → T (WOT), if Tn h, g → T h, g

for each fixed h, g in H . (a) Show that if S is the forward shift operator on 2 , then Sn → 0 (WOT), but Sn does not converge to 0 in either the norm or strong operator topology. (b) If Pn : 2 → 2 by Pn (x1 , x2 , . . .) = (0, 0, . . . , xn+1 , xn+2 , . . .) show that Pn → 0 (SOT), but not in the norm topology. (c) Show that if Tn → T (SOT) for Tn , T ∈ B(H ), then Tn → T (WOT). (d) Is the mapping from B(H ) → C which sends T to T  continuous if we use the strong operator topology (respectively, the weak operator topology) on B(H )?

Chapter 3

The Big Three

In linear spaces with a suitable topology, one encounters three far-reaching principles concerning continuous linear transformations. N. Dunford and J. Schwartz ([12], p. 49).

In this chapter we will look at several core results of functional analysis. Two of these, the principle of uniform boundedness and the open mapping theorem (as well as their close cousins, the Banach–Steinhaus theorem and the closed graph theorem) are Banach space results. The third, the Hahn–Banach theorem, makes no use of completeness and takes place in a normed linear space (or even more generally). All three of these results are ubiquitous in functional analysis. We will look at the Hahn–Banach theorem first, and begin by reviewing Zorn’s lemma, which is sometimes called the “analysts’ version of the axiom of choice.” We begin with some needed terminology. A partial order on a set X is a relation, written generically as ≤, satisfying the following properties for all a, b, c ∈ X: (a) (b) (c)

transitivity: if a ≤ b and b ≤ c then a ≤ c, reflexivity: a ≤ a, and anti-symmetry: if a ≤ b and b ≤ a then a = b.

If for every pair a, b in X we either have a ≤ b or b ≤ a, then X is said to be totally ordered by ≤. Two simple examples to illustrate these concepts are the totally ordered set consisting of the real line with the usual ≤ relationship, and, for any set X, any collection of subsets of X partially ordered by set inclusion ⊆. Note that the latter need not be a total ordering, for example when the cardinality of X is at least two, and P(X) is the set of all subsets of X, then (P(X), ⊆) is a partial ordering which is not a total ordering. As further examples, consider the set N × N of ordered pairs of positive integers and two relations ≤1 and ≤2 defined as follows. Say that (a, b) ≤1 (x, y) if a is strictly less than x or if a = x and b is less than or equal to y (in the usual sense on N). This is sometimes called lexicographical ordering, by its analogy with the ordering of words in a dictionary. For the second relation, say (a, b) ≤2 (x, y) if a is less than or equal to x, and b is less than or equal to y, in the usual sense of inequality on N. Clearly ≤1 is a total ordering, while ≤2 is a partial ordering which is not a total ordering.

B.D. MacCluer, Elementary Functional Analysis, DOI 10.1007/978-0-387-85529-5 3, c Springer Science+Business Media, LLC 2009 

49

50

3 The Big Three

When a set X is partially ordered by ≤ and Y is a subset of X, we call an element p ∈ X an upper bound for Y if y ≤ p for all y ∈ Y . An element m in X for which m ≤ x implies m = x is called a maximal element of X. We are ready to state Zorn’s lemma, which is indispensable in functional analysis. It will be taken as an axiom of set theory; alternatively one can derive it from the axiom of choice, which says that the Cartesian product of nonempty sets is nonempty. Lemma 3.1 (Zorn’s Lemma). If X is a nonempty partially ordered set with the property that every totally ordered subset of X has an upper bound in X, then X has a maximal element. A good way to see a standard and completely straightforward application of Zorn’s lemma is to use it to prove that any orthonormal set in a Hilbert space can be extended to an orthonormal basis, so in particular every Hilbert space has an orthonormal basis. See Exercise 3.1. The proof of the Hahn–Banach theorem in the next section will provide another example of an application of Zorn’s lemma.

3.1 The Hahn–Banach Theorem The Hahn–Banach theorem deals with extending continuous linear functionals from a subspace of a normed linear space to the whole space. Completeness of the space plays no role, so this is a result about normed linear spaces in general. This is one of the few places in the subject where it is helpful to first look at real vector spaces (that is, a vector space over R), and prove it in that context, before extending the argument to cover the case of a complex vector space. This extension, while not difficult, is not trivial either; historically the real case was proved nearly ten years before its extension to complex scalars. Theorem 3.2 (Hahn–Banach Theorem). Let X be a normed linear space over F = R or C, and suppose Y is a (not necessarily closed) proper subspace of X. If ϕ0 : Y → F is a bounded linear functional, then there is a bounded linear functional ϕ : X → F with the restriction of ϕ to Y equal to ϕ0 (so that ϕ is an extension of ϕ0 ) and ϕ  = ϕ0  (so that this extension is norm-preserving). It is the norm-preserving part of the conclusion that gives the result its power. Recall that Exercise 1.22 in Chapter 1 showed how to extend linear functionals on subspaces of Hilbert spaces, using simple Hilbert space techniques, in a normpreserving way. Without Hilbert space machinery at our disposal, we will have to work a bit harder. Before we look at the proof of the Hahn–Banach theorem, we give some applications. For the first, suppose X is a normed linear space which is not just {0}. Could X ∗ , the dual space of X, consist of just the zero functional? Corollary 3.3. Let X = {0} be a normed linear space. Given x0 = 0 in X, there is a bounded linear functional ϕ on X of norm 1 with ϕ (x0 ) = x0 .

3.1 The Hahn–Banach Theorem

51

Proof. Set M = {α x0 : α ∈ F}, a subspace of X. Define ϕ on M by ϕ (α x0 ) = α x0 . It is easy to see that ϕ is a bounded linear functional on M with norm 1 and ϕ (x0 ) = x0 . By the Hahn–Banach theorem we can extend ϕ to all of X without increasing its norm. The next corollary shows that the dual of a nontrivial normed linear space must be rich enough to “separate the points” of X. Corollary 3.4. Suppose X = {0} is a normed linear space. Given x1 = x2 we may find a bounded linear functional ϕ on X with ϕ (x1 ) = ϕ (x2 ). Proof. Apply Corollary 3.3 to x0 = x1 − x2 .



Corollary 3.5. Suppose x0 is an element of a normed linear space X. We have x0  = sup{|ϕ (x0 )| : ϕ ∈ X ∗ , ϕ  = 1}, and moreover this supremum is attained. Proof. The result if trivially true if x0 = 0. In general, we have |ϕ (x0 )| ≤ x0  if ϕ  = 1, so the supremum is at most x0 . On the other hand, by Corollary 3.3 there exists ϕˆ ∈ X ∗ with ϕˆ  = 1 and ϕˆ (x0 ) = x0 , so the supremum must in fact be equal to x0 , and the supremum is attained (by ϕˆ ). See Exercise 3.5 for a concrete application of Corollary 3.5 in the particular normed linear space L p (X, M, µ ), 1 ≤ p < ∞. Note the symmetry in the statements: For ϕ ∈ X ∗ , ϕ  = sup{|ϕ (x)| : x ∈ X, x = 1}, and for x ∈ X,

x = sup{|ϕ (x)| : ϕ ∈ X ∗ , ϕ  = 1}.

In the second line (but not in general in the first, although see Corollary 3.7 below), “sup” can be replaced by “max.” The Hahn–Banach theorem gives a means for determining the points that lie in the closure of a linear subspace of a normed linear space. The proof of the next result is left to the reader as Exercise 3.8. Corollary 3.6. Suppose that X is a normed linear space, x0 ∈ X and M is a (not necessarily closed) subspace in X. Suppose that d ≡ dist(x0 , M) > 0 where dist(x0 , M) = inf{x0 − y : y ∈ M}. There exists ϕ ∈ X ∗ with ϕ (x0 ) = 1, ϕ = 0 on M, and ϕ  = 1/d. In particular, x0 is in the closure of M if and only if there is no bounded linear functional on X that is 0 on M and nonzero at x0 . When X is a Banach space, or even just a normed linear space, we know from Exercise 2.1 in Chapter 2 that its dual X ∗ is a Banach space, so that it too has a dual space (X ∗ )∗ which we will write X ∗∗ . If x0 ∈ X we can define what will constitute

52

3 The Big Three

an element of X ∗∗ , call it x0∗∗ , by setting x0∗∗ (ϕ ) = ϕ (x0 ), for any ϕ in X ∗ . It is easy to see that x0∗∗ as just defined is a linear functional on X ∗ and moreover, |x0∗∗ (ϕ )| = |ϕ (x0 )| ≤ ϕ x0  so that x0∗∗ is bounded with norm at most x0 . We claim that this natural map from X to X ∗∗ sending x0 to x0∗∗ is a linear isometry of X into X ∗∗ . The interesting piece that still needs verification is the “isometry” part of this statement. We have x0∗∗  = sup{|x0∗∗ (ϕ )| : ϕ ∈ X ∗ , ϕ  = 1} = sup{|ϕ (x0 )| : ϕ ∈ X ∗ , ϕ  = 1} = x0 

where we have used the definition of the norm on X ∗∗ , the definition of x0∗∗ , and Corollary 3.5 for each of the three equalities, respectively. If this natural map x0 → x0∗∗ is onto X ∗∗ , then X is said to be reflexive, and it gives a isometric isomorphism of X with X ∗∗ . In this context, “isomorphism” means a continuous linear bijection with continuous inverse. If X is reflexive, it must be a Banach space since X ∗∗ is a Banach space. Note that the definition of “reflexive” requires that a particular mapping of X into X ∗∗ (the “natural map”) be an isometric isomorphism, not simply that there exist some isometric isomorphism of X and X ∗∗ . The latter surprisingly turns out to be a strictly weaker assumption; this was shown by R.C. James in [23]. As an immediate consequence of the definition of a reflexive space and Corollary 3.5 we see that bounded linear functionals on reflexive Banach spaces attain their norm; this is the content of the next result, whose proof is left to the reader. Corollary 3.7. Suppose that X is a reflexive Banach space. Given ϕ ∈ X ∗ , there exists a unit vector x0 in X such that |ϕ (x0 )| = ϕ . This corollary is sometimes useful for showing a particular Banach space is not reflexive; see, for example, Exercise 3.34. The converse to Corollary 3.7 is also true: In any nonreflexive Banach space there is a bounded linear functional which does not attain its norm on the unit sphere. This result is also due to James [24]. We are going to prove Theorem 3.2 in the real scalar case first, that is, we will assume in the statement of the theorem that F = R. There are two key parts to the proof: the “one-step extension,” which basically extends a linear functional, in a norm-preserving way, from a linear manifold to the span of that manifold and a single additional vector; and a Zorn’s lemma argument. Proof (Theorem 3.2, real case). If ϕ0  = 0, simply set ϕ ≡ 0 and we are done. Thus we are interested in the case ϕ0  = 0, and we may assume, without loss of generality, that ϕ0  = 1, which we do. Choose a vector z which lies in X but not in Y and let Y1 ≡ {y + α z : y ∈ Y and α ∈ R} = {α z − y : y ∈ Y and α ∈ R},

3.1 The Hahn–Banach Theorem

53

so that Y1 is the span of Y and z. For any choice of a fixed real number c, if we define ϕ1 on Y1 by ϕ1 (α z − y) = α c − ϕ0 (y), then ϕ1 is an extension of ϕ0 to a linear functional on Y1 . There are a few details to be checked here, starting with the observation that ϕ1 is well-defined (once c is chosen) and linear. Well-definedness follows from uniqueness of representation for every vector in Y1 in the form α z − y with y ∈ Y and α ∈ R. Once these easy issues are attended to (we leave the details to the reader), the issue becomes whether we can choose c so that ϕ1  = ϕ0  = 1. In other words, we want to choose c so that |α c − ϕ0 (y)| ≤ α z − y

(3.1)

for all y in Y and all scalars α in R. Now we want (3.1) to hold for all y ∈ Y and real scalars α , so it can be rewritten in an equivalent manner by dividing by |α | (since the α = 0 case is trivially true). Doing this, we see that the condition (3.1) we wish to satisfy is equivalent to the condition |c − ϕ0 (α −1 y)| ≤ z − α −1 y for all y ∈ Y and real α = 0, or more simply, to the condition |c − ϕ0 (y)| ≤ z − y

(3.2)

for all y ∈ Y . This last inequality will hold for some choice of c precisely when there is a choice of c satisfying

ϕ0 (y) − y − z ≤ c ≤ ϕ0 (y) + y − z

(3.3)

for all y ∈ Y . If we denote the left-hand side, ϕ0 (y) − y − z, by A(y) and the right-hand side, ϕ0 (y) + y − z, by B(y), a real c of the desired type can be found provided  [A(y), B(y)] y∈Y

is nonempty (and if it is nonempty, any c that lies in this intersection will do). We claim that  [A(y), B(y)] y∈Y

is nonempty precisely when A(y) ≤ B(v) for all y and v in Y . One direction of this claim is clear: If c ∈ [A(y), B(y)] for all y ∈ Y , then c ≥ A(y) for all y and c ≤ B(v) for all v. Conversely, if A(y) ≤ B(v) for all choices of y and v, then a ≡ supy∈Y A(y) ≤ infv∈Y B(v) ≡ b, and any c in the range a ≤ c ≤ b will lie in [A(y), B(y)] for all y. This verifies the claim, and we can create a norm-preserving extension if

ϕ0 (y) − y − z ≤ ϕ0 (v) + v − z for all y and v in Y . We have

(3.4)

54

3 The Big Three

ϕ0 (y) − ϕ0 (v) = ϕ0 (y − v) ≤ y − v ≤ y − z + z − v, where we have used the assumption that ϕ0 has norm 1. Rearranging this computation gives (3.4), and we conclude that there is a norm-preserving extension ϕ1 of ϕ0 from Y to Y1 . To finish the proof, we use Zorn’s lemma. Let P be the collection of all pairs (Y  , ϕ  ) where Y  is a (not necessarily closed) subspace containing Y and ϕ : Y  → R is a linear functional extending ϕ0 : Y → R with ϕ   = ϕ0  = 1. Partially order P by (Y  , ϕ  ) ≤ (Y  , ϕ  ) if Y  ⊆ Y  and the restriction of ϕ  to Y  is ϕ  . Suppose C ≡ {(Yβ , ϕβ ) : β ∈ I } is a totally ordered subset of P. Set N ≡ ∪β Yβ . Since C is totally ordered, N is a subspace. Define ϕ˜ on N by ϕ˜ (y) = ϕβ (y) if y ∈ Yβ ; note that ϕ˜ is well-defined since C is totally ordered, and is linear. Moreover, (N, ϕ˜ ) is in P; in particular there is an β so that |ϕ˜ (y)| = |ϕβ (y)| ≤ y so ϕ˜  ≤ 1. We see that (N, ϕ˜ ) is an upper bound for C , since (Yβ , ϕβ ) ≤ (N, ϕ˜ ) for all β . By Zorn’s lemma, P has a maximal element, which we denote (X∞ , ϕ∞ ). We must have X∞ = X, else we could do the one-step extension process to extend to the span of X∞ and x0 , where x0 is in X but not in X∞ , contradicting the maximality of (X∞ , ϕ∞ ). Once we know that X∞ = X, we have the desired norm-preserving exten sion ϕ∞ of ϕ0 to all of X. In Exercise 3.4, the reader can work through an application of the one-step extension process in a concrete setting. The extension of the proof of the Hahn–Banach theorem from the real case to the complex case is outlined in Exercises 3.2 and 3.3. While it is not hard, historically there was a span of nearly ten years between the work on the real case by Banach and the extension to the complex case by H. Bohnenblust and A. Sobczyk in 1938. Perhaps not coincidently, Banach’s esteemed 1932 treatise Op´erations Lin´eaires deals only with real Banach spaces. In the particular setting of X = L p the complex case appeared in 1936 in the work of F. Murray; see also the comment in Exercise 3.3 on an earlier contribution by H. L¨owig. The work of Bohenblust and Sobczyk may be the first place that the result is referred to as the “Hahn–Banach Theorem.” In reality, it would be more accurate to credit Eduard Helly with the first proof of a Hahn–Banach type theorem for work dating from 1912. Helly was working on a problem, posed by Riesz, which he reformulated as a problem about extending a bounded linear functional on a subspace of C[a, b]. His argument had a thoroughly modern flavor, and was quite similar to that later used independently by Hahn (1927) and Banach (1929). A key feature was the one-step extension process, and in particular the inequalities of (3.3) and (3.4) for the special case of C[a, b] appears in his work. A slightly later paper of Helly’s [22], published in 1921, gives the Hahn– Banach theorem in the context of sequence spaces. One possible explanation for the lack of recognition Helly received for this (and other mathematical contributions he made) might be found in some of the

3.2 Principle of Uniform Boundedness

55

non-mathematical details of his life (see [20]). An Austrian, he received his Ph.D. in 1907. He enlisted in the army at the start of World War I, and was wounded in 1915, suffering serious heart and lung injuries. He became a Russian prisoner of war, and was imprisoned in Siberia from 1915–1917. In part due to the civil war in Russia in 1918, he was not able to return to his home of Vienna until late in 1920. He received his Habilitation degree from the University of Vienna in 1921, but was unable to secure an academic position. He worked as a bank clerk (until the bank failed in 1929) and in an insurance company, while trying to remain active in mathematics. Helly— who was Jewish—emigrated to the United States in 1938, as Austria was absorbed by the Third Reich. He held positions at several junior colleges in New Jersey until he was offered a Professorship at the Illinois Institute of Technology. Unfortunately, shortly after this he died of a heart attack, at the age of 59. Undoubtedly his earlier war injuries contributed to this premature death.

3.2 Principle of Uniform Boundedness Several important problems in Banach space theory come down to the rather pedestrian-seeming problem of showing that a set is “large” in the sense that it has nonempty interior. Recall that the interior of a set A in a metric space (M, d) is the set of points a ∈ A for which there exists δ > 0 with B(a, δ ) ≡ {m ∈ M : d(a, m) < δ } ⊆ A. As an illustration of this general principle, let us characterize boundedness of a linear operator in terms of a particular set having nonempty interior. Proposition 3.8. Suppose X and Y are normed linear spaces and T : X → Y is linear. Then T is bounded if and only if T −1 ({y ∈ Y : y < 1}), the preimage of the open unit ball in Y under T , has nonempty interior. Proof. First suppose T is bounded with T  = M. If x ∈ B(0, 1/M) we have T x < 1, and thus B(0, 1/M) is contained in the preimage of the open unit ball of Y under T. The more interesting direction is the “if” direction. For this, suppose T is linear and T −1 {y : yY < 1} contains x0 as an interior point, say B(x0 , ε ) lies in the preimage of the open unit ball in Y under T . Fix ε  , 0 < ε  < ε , and consider x with x ≤ ε  . Since x + x0 ∈ B(x0 , ε ), we have T x = T (x + x0 − x0 ) = T (x + x0 ) − T (x0 ) ≤ T (x + x0 ) + T x0  ≤ 1 + T (x0 ) ≡ M. Thus for any unit vector v in X,

 

1 

≤ M T T v =

ε v

ε ε and T is bounded.



56

3 The Big Three

So when does a set have nonempty interior? In a complete metric space (thus in particular in a Banach space) the Baire category theorem sheds some light on this. Definition 3.9. A set S in a metric space M is nowhere dense if its closure has empty interior. Some examples of nowhere dense sets are the integers Z in the real line R, or the Cantor set in [0, 1] (it is a closed set containing no open intervals). By contrast, the rationals are not a nowhere dense set in R. The next result is the Baire category theorem. Its roots are in Ren´e Baire’s 1899 dissertation, where it was shown that Rn is not a countable union of nowhere dense sets (the case n = 1 was actually proved two years earlier by W. Osgood). One can think of this as a topological tool; it will play a role in the proof of the principle of uniform boundedness of this section, and the proof of the open mapping theorem in the next. Theorem 3.10 (Baire Category Theorem). A complete metric space is not the union of a countable number of nowhere dense sets. Proof. Let M be a complete metric space. Suppose, for a contradiction, that M is the countable union of sets An that are nowhere dense. We will construct a Cauchy sequence in M with no limit point in M. / where Since A1 is nowhere dense, we may find an open ball B1 with B1 ∩ A1 = 0, A1 denotes the closure of A1 . Clearly, we can choose the radius of this ball B1 to be less than 1. Since A2 has empty interior, it doesn’t contain B1 and (M\A2 ) ∩ B1 is a nonempty open set. Thus we may find an open ball B2 whose closure is contained in B1 such that B2 ∩ A2 = 0/ and such that the radius of B2 is less than 1/2. Continue inductively, so that at the nth stage we produce an open ball Bn whose closure is contained in Bn−1 , such that Bn ∩ An = 0/ and the radius of Bn is less than 1/n; we are using the hypothesis that An has no interior point and thus Bn−1 is not contained in An . Note that the closed balls Bn form a nested sequence of closed sets, whose diameters tend to 0, in our complete metric space. Let xn be the center of the ball Bn . It is easy to see that {xn } forms a Cauchy sequence in M: When m, n ≥ N, the points xm and xn lie in the ball BN and hence d(xm , xn ) < 2/N, which tends to 0 as N → ∞. This Cauchy sequence then converges to some point x ∈ M. We claim that x is not in A j for all j ≥ 1, a contradiction to the assumption that M = ∪∞j=1 A j . To verify this claim, note that if x ∈ A j , then x is not in B j (since B j ∩ A j = 0), / and hence not in Bk for all k ≥ j + 1. However, we have xn ∈ B j+1 for all n ≥ j + 1 and therefore x is in B j+1 ⊆ B j , a contradiction. Thus the Baire category theorem says that if a complete metric space X is written as a countable union X = ∪∞ 1 An , at least one of the sets An must be “big” in the sense that its closure has nonempty interior. The proof of the next result, the principle of uniform boundedness, illustrates this. Informally, this theorem says that a pointwise bounded family of operators is uniformly bounded.

3.2 Principle of Uniform Boundedness

57

Theorem 3.11 (Principle of Uniform Boundedness). Suppose X is a Banach space and F is a family of bounded linear operators from X to a normed linear space Y . If, for every x ∈ X, sup{T x : T ∈ F } < ∞ then sup{T  : T ∈ F } < ∞. Proof. Define An ≡ {x ∈ X : T x ≤ n for all T ∈ F }. The hypothesis says that each x ∈ X is in some An , so that X = ∪∞ n=1 An . By the Baire category theorem, for some n, An has nonempty interior. To make use of this we first claim that each An is closed. To see this, suppose xm is in An for m = 1, 2, . . . and that xm → x. For each T ∈ F , T xm  ≤ n, while continuity of T guarantees that T xm  → T x. Hence T x ≤ n for each T ∈ F and x ∈ An . So in fact, for some fixed n, An has an interior point, which we will denote x0 . Let ε > 0 be chosen so that B(x0 , ε ) ⊆ An . The positive number ε , as well as the integer n, are fixed values at this point. If x ≤ ε , then for any T ∈ F , T x = T (x + x0 ) − T x0  ≤ T (x + x0 ) + T x0  ≤ n + n. From this it follows that for any unit vector v, 1 2n T v = T (ε v) ≤ ε ε and thus sup{T  : T ∈ F } ≤

2n . ε



Note that we can restate the principle of uniform boundedness as follows: either sup{T  : T ∈ F } < ∞, or there exists x ∈ X such that sup{T x : T ∈ F } = ∞. As with the Hahn–Banach theorem, Helly deserves more credit than he has received for his contributions to the uniform boundedness principle. He gave the first proof, for C[a, b], but by methods which extend to general Banach spaces. Banach and Steinhaus’s original proof depended on a technique called the “gliding hump” method; this was replaced by the Baire category argument after S. Saks pointed out the possibility of using this approach. The original gliding hump argument is outlined in Exercise 3.15. A precursor to the principle of uniform boundedness, in the setting of 2 , appeared in work of E. Hellinger and O. Toeplitz in 1910. A close cousin of the principle of uniform boundedness is the Banach–Steinhaus theorem, which we look at next. Theorem 3.12 (Banach–Steinhaus Theorem). Suppose {Tn } is a sequence of bounded linear operators from a Banach space X to a Banach space Y . Assume further that for all x ∈ X, limn→∞ Tn x exists. Define T : X → Y by T x = limn→∞ Tn x. With this definition, T is a bounded linear operator from X to Y .

58

3 The Big Three

Proof. It is easy to check that T is linear, and we leave the details of this to the reader. We will use the principle of uniform boundedness to show that T is bounded. For each x ∈ X, supn Tn x < ∞ since {Tn x} is a convergent sequence by hypothesis and convergent sequences in a metric space are bounded. By Theorem 3.11, supn Tn  ≡ M < ∞. This means that for each x ∈ X, and any n, Tn x ≤ Mx. We have T x =  lim Tn x ≤ Mx n→∞

(note the continuity of the norm lurking behind this calculation) and T is bounded with T  ≤ M. Note that the last result does not say that if Tn → T pointwise on X, then Tn  → T . For example, let Tn : 2 → C be the linear operator given by Tn ({ak }) = an . For each a = {ak } in 2 we have limn→∞ Tn (a) = 0, so T ≡ 0 is the pointwise limit of the operators Tn . However, Tn − T  = Tn  = 1. We give some examples to illustrate applications of Theorems 3.11 and 3.12. The first uses the sequence space c0 = {{an }∞ 1 : limn→∞ an = 0}, in the supremum norm. It is left to the reader (see Exercise 3.16) to show that c0 is a closed subspace of ∞ , hence is itself a Banach space. Example 3.13. Suppose that {an }∞ 1 is a sequence of complex numbers such that ∞ ∞ ∑∞ 1 an bn converges whenever {bn }1 is in c0 . We will show that ∑1 |an | < ∞. To see this, define Tk : c0 → C by Tk ({bn }) =

k

∑ a jb j.

j=1

Each Tk is a bounded linear functional on c0 with Tk  ≤ ∑kj=1 |a j |; the latter statement follows from the calculation     k   k k k    ∑ a j b j  ≤ ∑ |a j b j | ≤ max |b j | ∑ |a j | ≤ {bn }∞ ∑ |a j |.  j=1  j=1 1≤ j≤k j=1 j=1 In fact, we have equality: Tk  = ∑kj=1 |a j |. To see this, consider Tk acting on the unit vector in c0   a1 a2 ak , ,···, ,0··· |a1 | |a2 | |ak | (with the obvious modifications if some a j = 0), whose image under Tk is ∑kj=1 |a j |. We are given, then, that for each b = {bn } ∈ c0 , limk→∞ Tk (b) exists, since Tk (b) = ∑k1 a j b j and ∑∞ 1 a j b j converges. In particular, supk |Tk (b)| < ∞ for each b ∈ c0 . By Theorem 3.11, supk {Tk } < ∞, and thus ∑∞j=1 |a j | < ∞. Example 3.14. Our next example is an application of Theorem 3.11 to a question about convergence of Fourier series. A continuous function f on the unit circle T has a Fourier series ∞



k=−∞

ak eikx ,

3.2 Principle of Uniform Boundedness

where

59

1 ak = fˆ(k) = 2π

 π −π

f (t)e−ikt dt

(that is, the Fourier expansion of f with respect to the orthonormal basis {einx }∞ −∞ for L2 (T, dx/(2π )). We know that this series converges to f in the norm of L2 (T ), and so a subsequence converges pointwise almost everywhere; this is true for any f in L2 (T ), not just the continuous ones. But the general question of the pointwise convergence of the series is more delicate. In fact, for a period of time that stretched for 40 years, Riemann, Dirichlet, Weierstrass, and Dedekind all believed that the Fourier series of a continuous function converged pointwise everywhere (necessarily to the function value). The first counterexample was given in 1876 by DuBois-Reymond. Then the pendulum swung and for a while it was believed that the Fourier series of a continuous function could fail to converge at every point. In 1966 Lennart Carleson settled the matter definitively by proving the “Lusin conjecture,” which asserts that the Fourier series of any function in L2 (T ) (and thus, in particular, of any continuous function) converges pointwise almost everywhere. What we will do in this example is use the principle of uniform boundedness to show there exists an f ∈ C(T ) such that sn ( f , 0) does not converge to f (0), where sn ( f , 0) denotes the symmetric partial sum ∑nk=−n fˆ(k)eikt evaluated at 0. Indeed, we will show the existence of an f ∈ C(T ) so that the partial sums of the Fourier series of f at t = 0 are unbounded. We begin with a calculation. sN ( f ,t) ≡

N



fˆ(k)eikt =

k=−N



N



k=−N

= =

 π

−π

 π

−π

π

−π

f (x)e−ikx

N

f (x)



dx 2π

eik(t−x)

k=−N

f (x)DN (t − x)

 eikt

dx 2π

dx 2π

where DN (s) = ∑Nk=−N eiks ; this is the so-called Dirichlet kernel. The reader may recognize the last integral as the convolution f ∗ DN (t) of f and DN . Next we claim that sin(N + 12 )s DN (s) = sin( 2s ) when s = 0, and 2N + 1 when s = 0. In the case s = 0 this is clear; otherwise write N



k=−N

2N

eiks = e−iNs ∑ eiks = e−iNs k=0

1 − ei(2N+1)s . 1 − eis

Multiplying numerator and denominator by e−is/2 and using the identity e−iy − eiy = −2i sin y gives the desired result. This kernel DN (s) is badly behaved in two respects: it is not positive and DN 1 is not bounded. To see the latter, we have

60

3 The Big Three

DN 1 =

 π | sin(N + 12 )s| ds



−π



| sin( 2s )|

=

 π | sin(N + 12 )s| ds 0

 2 π | sin(N + 12 )s|

| sin( 2s )|

π

ds π 0 s 1  2 π (N+ 2 ) du = | sin u| , π 0 u

where we have used the estimate 0 ≤ sint ≤ t for 0 ≤ t ≤ π /2 and made the substitution u = (N + (1/2))s. Now 2 π

 π (N+ 1 ) 2 0

| sin u|

2 N du ≥ ∑ u π k=1

 kπ (k−1)π

| sin u|

4 1 4 N 1 du = 2 ∑ ≥ 2 log(N + 1). kπ π k=1 k π

Now let us get set up to use the principle of uniform boundedness. Recalling that C(T ) is a Banach space in the supremum norm, define the linear functional Λn : C(T ) → C by Λn ( f ) = sn ( f , 0). Since sn ( f , 0) =

 π −π

f (x)Dn (−x)

dx 2π

we see that

 π  π  dx  dx  f (x)Dn (−x)  ≤ | f (x)||Dn (−x)| ≤  f ∞ Dn 1 |Λn ( f )| =  2 π 2 π −π −π

so that Λn is bounded with norm at most Dn 1 . We claim that we actually have equality: Λn  = Dn 1 . To see this, fix n and let g(x) be defined to be 1 if Dn (x) > 0, to be −1 if Dn (x) < 0 and 0 if Dn (x) = 0. We may then find continuous and piecewise linear functions f j (x) with −1 ≤ f j (x) ≤ 1 for all x and f j → g pointwise on [−π , π ] as j → ∞. By the dominated convergence theorem  π

lim

j→∞ −π

f j (x)Dn (−x)

dx = 2π

 π −π

g(x)Dn (−x)

dx = 2π

 π −π

|Dn (−x)|

dx = Dn 1 . 2π

Since  f j ∞ ≤ 1 this shows that Λn  ≥ Dn 1 as desired. We are finally ready to make our appeal to the principle of uniform boundedness. Either Λn  ≤ M for some M < ∞ and for all n, or there exists f ∈ C(T ) such that supn |Λn ( f )| = ∞. Since Λn  = Dn 1 → ∞, the first alternative cannot hold and thus the second must. We obtain the existence of an f ∈ C(T ) such that sup |Λn ( f )| = sup |sn ( f , 0)| = ∞, n

n

and the Fourier series of f diverges at 0. The next result is dual to the principle of uniform boundedness.

3.3 Open Mapping and Closed Graph Theorems

61

Theorem 3.15. Let X be a normed linear space, and suppose A is a subset of X. If sup{|ϕ (x)| : x ∈ A} is finite for each fixed ϕ in X ∗ , then A is bounded. Proof. Consider the natural map Φ : X → X ∗∗ taking x to x∗∗ . Note that Φ (A) is thus a collection of bounded linear functionals on X ∗ . Since X ∗ is a Banach space sup{|Φ (x)(ϕ )| : x ∈ A} = sup{|ϕ (x)| : x ∈ A} < ∞ for each ϕ ∈ X ∗ . By Theorem 3.11, applied to the linear maps F = {Φ (x) : x ∈ A}, we must have sup{Φ (x) : x ∈ A} < ∞. However, we know that Φ is an isometry of X into X ∗∗ , so that Φ (x) = x. Thus we conclude sup{x : x ∈ A} < ∞; that is, A is bounded. Exercise 3.18 gives an application of Theorem 3.15.

3.3 Open Mapping and Closed Graph Theorems The theorems of the title of this section are closely related; we will prove the open mapping theorem first, using the Baire category theorem, and then derive the closed graph theorem from it. An open map is one for which the image of every open set is open. The open mapping theorem concerns surjective maps in B(X,Y ). Theorem 3.16 (Open Mapping Theorem). Suppose that X and Y are Banach spaces and that T is a bounded linear operator from X to Y . If T maps X onto Y , then T (G) is open in Y whenever G is open in X. Before we discuss the proof, let us give one important consequence. This is often called the inverse mapping theorem, and it is the third member of the triumvirate of results in this section. Corollary 3.17 (Inverse Mapping Theorem). Suppose X and Y are Banach spaces and T ∈ B(X,Y ) is bijective. Its set-theoretic inverse T −1 is then a bounded linear operator from Y to X. Proof. We have already observed that T −1 exists as a linear map, so only boundedness of T −1 remains to be shown. By the open mapping theorem, T carries open sets to open sets. Now T −1 is bounded if and only if T −1 is continuous, and T −1 : Y → X is continuous if and only if (T −1 )−1 (G) is open in Y for every G that is open in X. But (T −1 )−1 (G) = T (G) and, by Theorem 3.16, T (G) is open in Y for any open set G in X. Thus we conclude that T −1 is bounded, as desired.

62

3 The Big Three

This answers our old question as to whether the set-theoretic invertibility of T ∈ B(X,Y ) implies its operator-theoretic invertibility, i.e., the existence of an inverse in B(Y, X). We see the answer is yes, a result that Paul Halmos calls “one of the pleasantest and most useful facts about operator theory.” The inverse mapping theorem was first proved by Banach in 1929. Our approach, using the open mapping theorem, is due to Schauder in 1930. We turn next to the proof of the open mapping theorem. We will accomplish this by first proving the next result. Theorem 3.18. Suppose that X and Y are Banach spaces, and let BX and BY denote the open unit balls, centered at 0, in X and Y , respectively. Suppose A is a bounded linear operator mapping X onto Y . There exists a positive constant δ such that δ BY ⊆ A(BX ); that is, given y ∈ Y with y < δ there is x ∈ X with x < 1 and Ax = y. Notice that the hypothesis that A is onto Y says that given any y in Y we may find an x in X with Ax = y; thus the significance of Theorem 3.18 is that we may control the norm of x in terms of the norm of y. Before we give the proof of Theorem 3.18, let us see that it will quickly yield the open mapping theorem.

Proof (Theorem 3.16). Let G be an open set in X and let x0 be in G. We only need to show that A(G) contains an open ball about Ax0 . To this end, translate G to obtain G ≡ G − x0 . Since G is an open set containing 0 we may find a positive number t with tBX ⊆ G . By Theorem 3.18 we have A(G ) ⊇ A(tBX ) = tA(BX ) ⊇ t δ BY for some positive constant δ . By linearity, A(G) = A(G + x0 ) = Ax0 + A(G ) ⊇ Ax0 + t δ BY ; this last is the open ball centered at Ax0 of radius t δ .



To prove Theorem 3.18 we first give a lemma which is an approximate version of the theorem. It says that given y ∈ Y we may get as close to y as desired by a vector of the form Ax for some x in X whose norm is controlled by the norm of y. Lemma 3.19. Suppose that X and Y are Banach spaces, and that A is a bounded linear operator mapping X onto Y . There is a positive number d with the following property: Given ε > 0 and y ∈ Y there exists x ∈ X such that Ax − y < ε and x < d −1 y. Proof. Given y ∈ Y there exists x˜ in X with Ax˜ = y, since A is surjective. This means Y=

∞  k=1

A(kBX )

3.3 Open Mapping and Closed Graph Theorems

63

where BX is the open unit ball in X. Since Y is a complete metric space, the Baire category theorem says that for some k, A(kBX ) has closure with nonempty interior; say A(kBX ) ⊇ B(y0 , r) for some r > 0 and y0 ∈ Y . If y < r, then y + y0 will be in B(y0 , r) and hence in A(kBX ). Thus for any y in Y with y < r we may find sequences {xn } and {xn } in kBX such that Axn → y0 and Axn → y0 + y. Consider xn ≡ xn − xn , and note that Axn → y and xn  < 2k. The conclusion will follow from exploiting linearity. Let z = 0 be an arbitrary vector in Y , so that (r/2)(z/z) is a vector in Y of norm less than r. By the first part of the proof we may find a sequence xn in X with xn  ≤ 2k and Axn → (r/2)(z/z). Linearity says A((2/r)zxn ) → z where the norm of (2/r)zxn is less than (4k/r)z. This is the desired conclusion, with d = r/(4k). We can now prove Theorem 3.18 by an iterative use of this lemma. In the statement of the Lemma 3.19, we will refer to y as the target vector and ε as the tolerance.

Proof (Theorem 3.18). Let A, X, and Y be as in the statement of the theorem, and let d be as given by Lemma 3.19. Fix y in dBY , the open unit ball of radius d centered at 0 in Y . We apply the lemma, with target vector y and tolerance ε = d/2, to find x1 ∈ X of norm less than (1/d)y < 1 such that y − Ax1  < d/2. Apply the lemma again, this time with target y − Ax1 and tolerance ε = d/4 to find x2 in X with (y − Ax1 ) − Ax2 
n, so that A − An  = sup j>n |α j | → 0 as n → ∞. This shows that A is the limit of finite rank operators, hence A is compact, by Theorem 4.10. This example also shows that the finite rank operators form a proper subclass of the compact operators. It is also the case that a compact diagonal operator must have its diagonal converging to 0; see Exercise 4.8. Although the result is true more generally, we will find it convenient in the next result to restrict our attention to Hilbert spaces with a countable orthonormal basis. Having a countable orthonormal basis is equivalent to being a separable Hilbert space, that is, having a countable dense subset; see Exercise 1.23 in Chapter 1. Nonseparable Hilbert spaces were not studied before 1934, and we will restrict our attention to the separable case whenever convenient.

4.2 Compact Operators

83

Theorem 4.11. If T ∈ B(H ) is a compact operator on a Hilbert space H having a countable orthonormal basis, there exists finite rank operators Tn with T −Tn  → 0; that is, every compact operator is a limit of finite rank operators. Proof. There are three steps to the proof: Defining some likely candidates for the operators Tn , showing that for each fixed h in H , Tn h → T h, and then using this pointwise convergence to show Tn − T  → 0. For the first step, suppose that {e1 , e2 , . . .} is an orthonormal basis for the closure of the range of T (a closed subspace of H ); notice it is only interesting if this basis is infinite. Let Pn be the projection onto the span of the first n vectors e1 , e2 , . . . , en . Notice that we are using here the fact that this span is a closed subspace of H , which follows from Proposition 4.3 or Exercise 1.21 in Chapter 1. Define Tn = Pn T , so that Tn is clearly a finite rank operator. Now we check convergence of the Tn to T at each point of H . Let h be any vector in H and set k = T h. We have n

Tn h = Pn T h = Pn k =

∑ k, e j e j ,

j=1

where we are using the result of Exercise 1.21 in Chapter 1, and ∞

Th =

∑ k, e j e j ,

j=1

since {e j } is an orthonormal basis for the closure of the range of T . Thus ∞

Tn h − T h2 =



| k, e j |2 ,

j=n+1

∑∞j=1 | k, e j |2

which tends to 0 as n → ∞, since = T h2 < ∞. Having established this pointwise convergence of Tn to T , we now consider Tn − T . Let Bc denote the closed unit ball in H . Since T is a compact operator, the closure of T (Bc ) is compact. Given ε > 0 the collection of open balls B(T h, ε ), centered at points T h for h ∈ Bc and with radius ε , forms an open cover of the closure of T (Bc ). By compactness, we may select a finite subcover, say T (Bc ) ⊆

m 

B(T h j , ε ),

(4.2)

j=1

for some positive integer m and some h j ∈ Bc . By our pointwise estimate, for each j, 1 ≤ j ≤ m, there is an integer N( j) so that Tn h j − T h j  < ε if n ≥ N( j). Set N∗ = max1≤ j≤m N( j) and let h be an arbitrary unit vector in H . By condition (4.2) we may find a value of j, 1 ≤ j ≤ m, so that T h − T h j  < ε . For any n ≥ N∗ we have

84

4 Compact Operators

Tn h − T h ≤ Tn h − Tn h j  + Tn h j − T h j  + T h j − T h = Pn T (h − h j ) + Tn h j − T h j  + T (h − h j ) ≤ 2T (h − h j ) + ε ≤ 3ε , where we have used the fact that the projection Pn has norm 1. Since h was an arbitrary unit vector, this calculation shows that if n ≥ N∗ , then Tn − T  ≤ 3ε , and since ε is arbitrary, this shows that Tn − T  → 0 as n → ∞, as desired. This theorem is true even without the hypothesis that H has a countable basis, that is, even for nonseparable Hilbert spaces; the necessary additions to the proof for this case can be found in [8]. It makes sense to ask what happens if H is replaced by a Banach space X, i.e., is every compact operator in B(X) a limit of finite rank operators? This question, known as the approximation problem, was formulated by T. Hildebrandt in 1931, and has been a problem of fundamental importance in Banach space theory. Over time, many equivalent properties were discovered, and various related approximation properties were defined. In 1973, Per Enflo caused a sensation by constructing a counterexample to the original approximation problem. Enflo’s work on this problem also yielded a negative answer to another long-open problem in functional analysis: Does every separable Banach space have a Schauder basis, that is, in any Banach space X is there always a sequence {x j } such that each x ∈ X can be uniquely written as   x=



n

j=1

j=1

lim ∑ c j x j ∑ c j x j = n→∞

?

It is true that in a Banach space with a Schauder basis, every compact operator is a limit of finite rank operators, and this includes all of the familiar Banach spaces. Enflo’s work provided a solution to Problem #153 in the “Scottish Problem Book”, posed by S. Mazur in 1936, for which a prize of “a live goose” had been offered1 . When Enflo was lecturing in Warsaw in 1972, Mazur presented him with the prize goose promised 36 years earlier. The next result is sometimes phrased as “K (H ) is self-adjoint” when H is a Hilbert space. Proposition 4.12. If T is in B(H ) for a separable Hilbert space H , then T is compact if and only if T ∗ is compact. Proof. Since T ∗∗ = T , it suffices to prove that T compact implies T ∗ is compact. By Theorem 4.11, if T is compact, there are finite rank operators Tn that converge to T . Now Tn − T  = Tn∗ − T ∗  by Proposition 2.14, and we will be done by an appeal to Theorem 4.10 if we can show that each Tn∗ is finite rank. Let Pn be the projection of H onto the range of Tn , a closed subspace of H . Each Pn is finite rank since each Tn is. Since Pn is a projection, Pn Tn = Tn , and taking adjoints, Tn∗ Pn∗ = Tn∗ . But 1

See Section 3.5.

4.2 Compact Operators

85

projections are self-adjoint, so we have Tn∗ = Tn∗ Pn . It is easy to see that the finite rank operators form a (two-sided) ideal (see Exercise 4.6), so we conclude that Tn∗ is finite rank, and thus T ∗ is compact. Since Theorem 4.11 extends to nonseparable Hilbert spaces, so does Proposition 4.12. The implication “T compact implies T ∗ compact” also extends to the Banach space setting, where it sometimes goes by the name of “Schauder’s Theorem”; see [8]. To explore a large class of compact operators, we give a definition. Definition 4.13. Suppose that H is a Hilbert space with a countable orthonormal basis, and let T be in B(H ). We say that T is Hilbert–Schmidt if there is an or∞ 2 thonormal basis {en }∞ 1 of H such that ∑n=1 Ten  < ∞. It is convenient to know that the sum appearing in this definition is actually independent of the choice of basis; this is the next result, which is due to von Neumann. Proposition 4.14. Suppose that T is a bounded linear operator on a separable Hilbert space H and {en }∞ n=1 is an orthonormal basis for H such that ∞

∑ Ten 2 < ∞.

n=1

For any other orthonormal basis { fn }∞ n=1 , we have ∞



n=1

n=1

∑ T fn 2 = ∑ Ten 2 .

Proof. The proof relies on repeated applications of Parseval’s identity. For each n we have ∞

T fn =

∑ T fn , f j f j and T fn 2 =

j=1



∑ | T fn , f j |2 .

(4.3)

j=1

Similarly, for each n and each j, Ten 2 =



∑ | Ten , f j |2 ,

(4.4)

j=1

T ∗ f j 2 =



∑ | T ∗ f j , en |2 ,

(4.5)

n=1

T ∗ f j 2 =



∑ | T ∗ f j , fn |2 .

(4.6)

n=1

Thus ∞











∑ T fn 2 = ∑ ∑ | T fn , f j |2 = ∑ ∑ | T ∗ f j , fn |2 = ∑ T ∗ f j 2 ,

n=1

n=1 j=1

j=1 n=1

j=1

86

4 Compact Operators

where we have used Equations (4.3) and (4.6). Similarly, by Equations (4.4) and (4.5) ∞











∑ Ten 2 = ∑ ∑ | Ten , f j |2 = ∑ ∑ | T ∗ f j , en |2 = ∑ T ∗ f j 2

n=1

n=1 j=1

so that

j=1 n=1





n=1

n=1

j=1

∑ T fn 2 = ∑ Ten 2 ,

as desired.

As a corollary of the proof of the last result, we see that if T is Hilbert–Schmidt, then so is T ∗ . The next result shows that a Hilbert–Schmidt operator is compact. Theorem 4.15. Every Hilbert–Schmidt operator on a separable Hilbert space is compact. Proof. Let A be Hilbert–Schmidt on H . We will exhibit A as a limit of finite rank operators and use Theorem 4.10 to conclude A is compact. Fix an orthonormal basis ∞ 2 {ek }∞ 1 of H , so that ∑k=1 Aek  < ∞. For each n ≥ 1, define An by   ∞ n ˆ ˆ An h = An ∑ h(k)e k = ∑ h(k)Aek , k=1

k=1

ˆ where h(k) = h, ek . Clearly An is linear, and An is finite rank, since the range of An is contained in the span of the vectors Ae1 , Ae2 , . . . , Aen . Moreover, for any h ∈ H , (A − An )h = 









ˆ h(k)Ae k ≤

k=n+1

so that

1  2

ˆ |h(k)|

2

k=n+1

 A − An  ≤











1 2

Aek 

2

k=n+1

1 2

Aek 

2

,

k=n+1

which tends to 0 as n → ∞.



The next result gives a large class of examples of Hilbert–Schmidt operators, and explains how they arise naturally. It concerns operators on L2 (X, µ ) for some measure space (X, µ ). Theorem 4.16. Suppose that L2 (X, µ ) is a separable Hilbert space and K is an integral operator on L2 (X, µ ), with kernel k(x, y) ∈ L2 (X × X). The operator K is Hilbert–Schmidt. 2 Proof. Let {en }∞ n=1 be a basis for L (X, µ ). For fixed x ∈ X, write kx (y) = k(x, y); 2 then kx is in L (µ ) for almost every x and we have

4.3 A Preliminary Spectral Theorem

87



(Ken )(x) =

X

= X

k(x, y)en (y)d µ (y) kx (y)en (y)d µ (y)

= kx , en , so that Ken 2 = and



∑ Ken 2 =

n=1







X

|Ken (x)|2 d µ (x) =



n=1 X



| kx , en |2 d µ (x) =

| kx , en |2 d µ (x)

X





∑ | kx , en |2 d µ (x).

X n=1

2 Now it is easy to see that {en }∞ n=1 is also an orthonormal basis for L (X, µ ), so that kx , en are the Fourier coefficients of kx with respect to this basis. By Parseval’s identity, ∞

∑ | kx , en |2 = kx 2 .

n=1

Thus we have ∞

∑ Ken 2 =

n=1

 X

kx 2 d µ (x)

 

= X



= X×X

X

 |k(x, y)|2 d µ (y) d µ (x)

|k(x, y)|2 d(µ × µ ) < ∞

since k ∈ L2 (µ × µ ).



4.3 A Preliminary Spectral Theorem Recall from linear algebra, that if M is a self-adjoint n × n matrix (meaning it is equal to its conjugate transpose M ∗ ) then all the eigenvalues of M are real, and there is a unitary matrix U (meaning U ∗ = U −1 ) such that UMU −1 is real and diagonal. There is an orthonormal basis for Cn consisting of eigenvectors of M. In fact, if M is normal, that is, if M ∗ M = MM ∗ , then M is unitarily diagonalizable and Cn has an orthonormal basis of eigenvectors. Results that are generalizations of this to operators on Hilbert spaces are called “spectral theorems.” They appear with several quite distinct-looking formulations, and their connection to the finite-dimensional linear algebra results can seem somewhat obscured. In this section we will obtain a spectral theorem for compact self-adjoint operators on a Hilbert space; one can view this as a complete description of such operators. Later, in Chapter 6, we will obtain a much more general spectral theorem

88

4 Compact Operators

for bounded normal operators, at which point we will revisit the main results of this section. Our first goal is to obtain some information about the eigenvalues of compact self-adjoint operators in B(H ). The notion of an eigenvalue of an operator in B(H ) is the expected one, and the definition looks the same for operators on Hilbert or Banach spaces. Definition 4.17. We say that λ ∈ C is an eigenvalue of T ∈ B(X), where X is a Banach space, if there is a nonzero vector x in X so that T x = λ x. When λ is an eigenvalue of T ∈ B(X), the kernel of T − λ I is called the eigenspace corresponding to the eigenvalue λ , and the nonzero vectors in the eigenspace are called eigenvectors. For a compact operator T with nonzero eigenvalue λ , the kernel of T − λ I is necessarily finite-dimensional; see Exercise 4.10 and Exercise 4.25. We will show that a compact self-adjoint operator T always has either T  or −T  as an eigenvalue. Before proceeding to the proof of this, we need a lemma. The role of self-adjointness in it, and in the next several results, is contained in the following observation: If T is self-adjoint, then for any vectors x and y, T x, y = x, Ty = Ty, x , and so in particular T z, z must be real for all vectors z. For any operator T in B(H ), it is easy to see that T  = sup{| T x, y | : x = 1, y = 1}. The next result refines this for self-adjoint operators. Lemma 4.18. Suppose T is a self-adjoint operator in B(H ) for some Hilbert space H . We have T  = sup | T x, x | . x=1

Proof. Set M = supx=1 | T x, x |. Our goal is to show M = T . We make three easy observations (a) (b) (c)

h h For each h = 0 ∈ H , | T h, h | = | T (h h ), h h

| ≤ Mh2 .

Since | T h, h | ≤ T hh ≤ T h2 we must have M ≤ T , by the definition of M. For all f , g ∈ H , T ( f + g), f + g − T ( f − g), f − g = 4Re T f , g . This follows from expanding T ( f ± g), f ± g , using the self-adjointness of T to write T f , g + T g, f = T f , g + T f , g = 2Re T f , g .

Since T is self-adjoint, T x, x and Ty, y are real for any x, y ∈ H , and we have

4.3 A Preliminary Spectral Theorem

89

T x, x − Ty, y ≤ | T x, x | + | Ty, y |

(4.7)

≤ M(x + y ) M = (x − y2 + x + y2 ), 2 2

2

where we have used the observation in (a) and the parallelogram equality. Now let v be any unit vector and suppose T v = 0. Let s = T v and put 1 1 x = v + T v and y = v − T v s s in Equation (4.7). We obtain M T x, x − Ty, y ≤ 2





2 2

T v + 2v2 = 4M.

s

Since by the calculation in (c) we have 1 T x, x − Ty, y = 4Re T v, T v = 4T v s we conclude that T v ≤ M for any unit vector v and hence T  ≤ M. Since we had already observed the reverse inequality, we are done. Theorem 4.19. If T is a compact self-adjoint operator in B(H ), then at least one of the numbers T  and −T  is an eigenvalue of T . Proof. Without loss of generality we assume T  = 0, else T = 0 and 0 is trivially an eigenvalue of T . By Lemma 4.18 we have T  = supx=1 | T x, x |. Find unit vectors xn with | T xn , xn | → T . Since T is self-adjoint, each T xn , xn is real, and passing to a subsequence if necessary (which we don’t relabel) we may assume that T xn , xn → λ where either λ = T  or λ = −T . Since λ is real, we have T xn − λ xn 2 = T xn 2 − 2Re T xn , λ xn + |λ |2 xn 2 = T xn 2 − 2λ T xn , xn + λ 2 ≤ T 2 − 2λ T xn , xn + λ 2 = 2λ 2 − 2λ T xn , xn . As n → ∞, 2λ 2 − 2λ T xn , xn → 0, so that T xn − λ xn → 0.

(4.8)

Since T is compact, {T xn } has a convergent subsequence, say T xnk → y. By (4.8), we have λ xnk → y, and thus λ T xnk → Ty. But λ T xnk → λ y, so that Ty = λ y. Are we done? Yes, if we can show that y = 0. We have T xnk  = λ xnk + T xnk − λ xnk 

90

4 Compact Operators

≥ λ xnk  − T xnk − λ xnk  = |λ | − T xnk − λ xnk , where we know that T xnk − λ xnk  → 0 and λ = 0. But T xnk → y, so that y = limn→∞ T xnk  = 0. The next result applies to any self-adjoint operator on a Hilbert space, compact or not. Theorem 4.20. Suppose that T is self-adjoint in B(H ). Every eigenvalue of T is real, and the eigenvectors for distinct eigenvalues are orthogonal. Proof. Suppose that for some nonzero vector h and some scalar λ , T h = λ h. Since T h, h = λ h, h = λ h2 and T h, h = h, T h = h, λ h = λ h2 we must have λ = λ , and λ is real. If λ and µ are distinct (real) eigenvalues for T with T h = λ h and T g = µ g then 0 = T h, g − h, T g = λ h, g − µ h, g = (λ − µ ) h, g and h, g = 0. The reader is cautioned that this result does not say that a self-adjoint operator must have eigenvalues. In contrast to the conclusion of Theorem 4.19, neither compactness of T nor self-adjointness of T is (separately) sufficient to guarantee the existence of an eigenvalue for T . Exercises 4.14 and 4.15 ask you to verify this assertion. Theorem 4.21. Suppose that T is a compact self-adjoint operator in B(H ). The set of eigenvalues of T is a finite or countably infinite set of real numbers; if infinite, the eigenvalues form a sequence that converges to zero. Proof. By Theorem 4.20, all the eigenvalues are real. Also observe that if T x = λ x, then |λ | ≤ T , so that no eigenvalue has absolute value greater than T . There is nothing further to do if the set of eigenvalues is finite, so suppose it is infinite. We claim that for each ε > 0 , there are at most finitely many eigenvalues with absolute value at least ε . Suppose this is not the case. We may then find a sequence {λ j } of distinct eigenvalues, with |λ j | ≥ ε and unit eigenvectors y j with Ty j = λ j y j . By Theorem 4.20 the y j are orthogonal, and thus Ty j − Tyk 2 = λ j y j − λk yk 2 = |λ j |2 + |λk |2 ≥ 2ε 2 . This is a contradiction, since the compactness of T guarantees that {Ty j } has a convergent subsequence. Thus the claim is verified and we have shown: (a) (b)

The set of eigenvalues is countable, since {λ : λ is an eigenvalue and |λ | ≥ 1/n} is finite for every positive integer n. If the eigenvalues are a countably infinite set, they form a sequence which converges to zero.

This completes the proof.



4.3 A Preliminary Spectral Theorem

91

A version of this result holds for arbitrary compact operators on a Hilbert or Banach space. The eigenvalues need not be real, but if infinite they still form a sequence converging to zero. In the next result, we write T M for {T m : m ∈ M}. Lemma 4.22. Suppose that T is a bounded operator on a Hilbert space H and that M is a closed subspace of H . If T M ⊆ M, then T ∗ M ⊥ ⊆ M ⊥ . Conversely, if T ∗ M ⊥ ⊆ M ⊥ , then T M ⊆ M. Proof. Since T ∗∗ = T and (M ⊥ )⊥ = M, only the first assertion needs to be verified. Let n be in M ⊥ and let m be in M. We must show that T ∗ n ⊥ m, or equivalently, T ∗ n, m = 0. We have T ∗ n, m = n, T m = 0

since T m is in M. The next corollary is immediate.

Corollary 4.23. If T is a self-adjoint operator in B(H ), and if T M ⊆ M for some closed subspace M, then T M ⊥ ⊆ M ⊥ . A closed subspace M is called an invariant subspace for T ∈ B(H ) if T M ⊆ M. It is called a reducing subspace for T if both T M ⊆ M and T M ⊥ ⊆ M ⊥ . By virtue of Lemma 4.22, M is reducing if and only if it is invariant for both T and T ∗ . For an easy example, note that the subspaces NE = { f ∈ L2 [0, 1] : f (x) = 0 almost everywhere on E} for any measurable subset E of [0, 1] are reducing subspaces for the operator Mx of multiplication by ϕ (x) = x on (L2 [0, 1], dx). On the other hand, N = { f ∈ La2 (D) : f (0) = 0} is an invariant subspace for the multiplication operator Mz on La2 (D), but it is not a reducing subspace since, for example, Mz∗ (z) is the constant 1/2, which is not in N. Further examples can be found in Exercise 4.17. The terminology “reducing subspace” is suggestive of how these subspaces are used: Since H = M ⊕ M ⊥ , if M is a reducing subspace of T then the study of T on H is “reduced” to its study on the (smaller) Hilbert spaces M and M ⊥ . We’ll see this in action in the next result, which is the spectral theorem for compact self-adjoint operators. Our presentation follows that in [48]. Theorem 4.24 (Spectral Theorem, Preliminary Version). Let T = 0 be a compact, self-adjoint operator in B(H ). There exists a finite or countably infinite orthonormal set {gn } of eigenvectors of T , with corresponding real eigenvalues {λn }, such that T x = ∑ λn x, gn gn . n

If {λn } is infinite, then λn → 0.

92

4 Compact Operators

Proof. The fact that all eigenvalues are real has already been shown. For the rest we give an inductive construction. We know that λ1 ≡ T  or −T  is an eigenvalue of T . Pick a corresponding unit eigenvector g1 . Let M1 be the span of {g1 }. Since T (α g1 ) = αλ1 g1 , M1 is an invariant subspace for T . By Corollary 4.23, it is a reducing subspace, i.e., T M1⊥ ⊆ M1⊥ . Let T2 be the restriction of T to the Hilbert space M1⊥ ≡ H2 . Now, the restriction of a compact operator to an invariant subspace is compact (see Exercise 4.21), so T2 ∈ B(H2 ) is compact. We claim that T2 is also self-adjoint: if x, y are in M1⊥ , then T2∗ x, y = x, T2 y = x, Ty = T x, y = T2 x, y , where the fact that T is self-adjoint is used. Applying Theorem 4.19 again, this time to T2 ∈ B(H2 ), we see that T2 has an eigenvalue λ2 with λ2 = T2  or −T2 , and corresponding unit eigenvector g2 ∈ H2 . Notice that |λ2 | ≤ |λ1 |, and of course, λ2 is also an eigenvalue of the original operator T . Since g2 is in H2 = M1⊥ , g2 ⊥ g1 . Proceed inductively: Suppose we have obtained pairwise orthogonal unit eigenvectors g1 , g2 , . . . , gn of T corresponding to real eigenvalues λ1 , λ2 , . . . , λn with |λ j | = T j , where T1 = T , and T j is the restriction of T to [span {g1 , g2 , . . . , g j−1 }]⊥ for 1 < j ≤ n. Let Tn+1 be the restriction of T to Hn+1 ≡ [span {g1 , g2 , . . . , gn }]⊥ . Since span {g1 , g2 , . . . , gn } is invariant under T , so is Hn+1 , by Corollary 4.23. Moreover, as above, Tn+1 : Hn+1 → Hn+1 is compact and self-adjoint, and thus Tn+1 must have eigenvalue λn+1 , equal to either Tn+1  or −Tn+1 , and we choose a corresponding unit eigenvector gn+1 ∈ Hn+1 ; this is of course also an eigenvector for T on H . By the definition of Hn+1 , gn+1 is orthogonal to g j for 1 ≤ j ≤ n. As we continue this process, one of two things will happen. Either there is a smallest m with Tm = 0, in which case the process terminates with the construction of gm−1 , or Tn = 0 for all n. In the first case, consider, for arbitrary x in H , y ≡ x−

m−1

∑ x, g j g j .

j=1

x, g j g j is the projection of x onto span {g1 , g2 , . . . , gm−1 }, the projecSince ∑m−1 1 tion theorem says that y is in the orthogonal complement of this span, that is, y is in Hm . Thus we have 0 = Tm y = Ty = T x − which says that

m−1

m−1

j=1

j=1

∑ x, g j T g j = T x − ∑ λ j x, g j g j ,

4.3 A Preliminary Spectral Theorem

93 m−1

Tx =

∑ λ j x, g j g j

j=1

for every x in H , giving the desired conclusion in this case. In the case that Tn is not zero for any n, note while there may be repeated values in the sequence λ1 , λ2 , . . ., by Exercise 4.10 each value appears only finitely many times. This observation, together with Theorem 4.21, says λn → 0. Again consider an arbitrary x in H . We wish to show ∞

Tx =

∑ λ j x, g j g j ,

j=1

i.e., that n−1

T x = lim

n→∞

∑ λ j x, g j g j .

j=1

To this end, set n−1

yn = x − ∑ x, g j g j j=1

and notice that yn is in Hn and ∑n−1 1 x, g j g j is in its orthogonal complement. In particular, this guarantees by the Pythagorean theorem that x ≥ yn . Now Tyn  = Tn yn  ≤ Tn yn  = |λn |yn  ≤ |λn |x. We know that |λn | → 0 as n → ∞ so we must have n−1

T x − ∑ λ j x, g j g j  = Tyn  ≤ |λn |x → 0, j=1

which is our desired conclusion.



We note that the λn appearing in the statement of the last result must form a complete list of all the nonzero eigenvalues of T . To see this, observe that if T z = µ z for some µ distinct from all the λn , then z ⊥ gn for all n, since the eigenvectors corresponding to distinct eigenvalues are necessarily orthogonal for any self-adjoint operator (Theorem 4.20). Hence

∑ λn z, gn gn = 0 = T z = µ z and z = 0. In the last result, the gn are an orthonormal sequence, but need not be an orthonormal basis for H . The next result explores this further. Corollary 4.25. If T is a compact self-adjoint operator on a separable Hilbert space H , then there is an orthonormal basis {en } of H consisting of eigenvectors for T such that

94

4 Compact Operators

T x = ∑ λn x, en en n

for every x in H , where λn is the eigenvalue of T corresponding to the eigenvector en . This sum is either finite or countably infinite. Proof. By Theorem 4.24, there is a finite or infinite orthonormal sequence {gn } such that (4.9) T x = ∑ λn x, gn gn n

and T gn = λn gn . By the construction in Theorem 4.24, the λn are nonzero. Let {hm } be an orthonormal basis for ker T ; this is at most countable since H is assumed to be separable. We have T hm = 0 and each hm is an eigenvector of T . Since eigenvectors corresponding to distinct eigenvalues are orthogonal, hm ⊥ gn for all m, n. Thus {gn } ∪ {hm } is an orthonormal set in H consisting of eigenvectors of T . We claim it is an orthonormal basis for H . By Equation (4.9), x − ∑ x, gn gn n

is in ker T so that

x − ∑ x, gn gn = ∑ cm hm n

m

for some coefficients cm ; in fact we must have cm = x − ∑ x, gn gn , hm = x, hm

n

since hm ⊥ gn . We have shown that an arbitrary x ∈ H can be written as x = ∑ x, gn gn + ∑ x, hm hm . n

m

Thus {gn }∪{hm } is a countable orthonormal basis for H . If we relabel this as {en }, we are done.

4.4 The Invariant Subspace Problem The invariant subspace problem, which has been variously described as “the most fundamental question in operator theory” [3] or “the most famous unsolved problem in the theory of bounded linear operators” [35], asks whether every T ∈ B(X) has a nontrivial closed invariant subspace; in this section the term “invariant subspace” will always mean a nontrivial closed invariant subspace. One can ask this question when X is a Banach space or when X is a Hilbert space and this distinction is important. Of course, if T has an eigenvalue, then the corresponding eigenspace is invariant for T , but it is easy to give examples of operators with no eigenvalues but yet having invariant subspaces (see Exercises 4.14 and 4.15).

4.4 The Invariant Subspace Problem

95

In 1950, rediscovering unpublished work of von Neumann, Nachman Aronszajn showed that a compact operator on a Hilbert space always has an invariant subspace. A few years later Aronszajn and Kennan Smith generalized this result to compact operators on Banach spaces. Paul Halmos described the situation subsequent to this work as follows [18]: Smith pointed out, I might almost say complained, that the proof was “tight”. It left no room for modifications and generalizations; it proved exactly what it was designed to prove, no more.... Aronszajn taught me the proof on a restaurant napkin several months before the paper appeared. I understood it, I cherished it, and along with many others I kept trying to “loosen” it so as to be able to apply it more broadly—but all to no avail (p. 320).

There the matter stayed until a breakthrough occured in 1966, and it was shown that any operator T on a Hilbert space H for which there is a nonzero polynomial p(z) = an zn + · · · + a1 z + a0 such that p(T ) = an T n + · · · + an T + a0 I is compact (such an operator T is said to be polynomially compact) has an invariant subspace. The first proof of this, by Allen Bernstein and Abraham Robinson, used methods of “nonstandard analysis,” but Halmos quickly reworked their argument to formulate them in classical standard analysis, publishing the resulting work as a short paper later the same year. In 1973, the young Russian mathematician Victor Lomonosov caused a sensation by announcing a theorem which included the following result: Any operator on an infinite-dimensional complex Banach space which commutes with a nonzero compact operator has an invariant subspace. Even more, he shows that any operator which commutes with an operator (not a scalar multiple of the identity) which commutes with a nonzero compact operator has an invariant subspace. At first it was not clear whether this latter description might include all bounded linear operators. While Lomonosov’s proof was short and elegant, an even briefer and more accessible proof was later provided by Hugh Hilden; the reader can find the details of Hilden’s argument in [43], pp. 120–121. This particular thread of work on the invariant subspace problem—starting with von Neumann and culminating with Lomonosov and Hilden—proceeds from the philosophy that compact operators generalize finite-dimensional operators. Another thread in the invariant subspace story is anchored by the statement that normal operators on a Hilbert space always have invariant subspaces (we’ll see this in Chapter 6). One then tries to find other classes of operators, related in some way to a weakening of the normality hypothesis, which can be shown to have invariant subspaces. In particular, the class of subnormal operators (see [5]) on Hilbert spaces all have invariant subspaces. As of this writing, the invariant subspace problem is still open for bounded linear operators on Hilbert spaces. For Banach space, though, the situation was resolved by work of Enflo published in 1987. He constructs a Banach space X and an operator in B(X) with no invariant subspace. While the date of publication is 1987, Enflo’s announcement of the result, and a manuscript containing the example, dates from 1975. Certainly part of the explanation for the long delay before formal publication lies in the complexity of Enflo’s construction. In fact, in reviewing the 100-page long Acta Mathematica publication, the reviewer A.M. Davie writes [9]:

96

4 Compact Operators [Enflo’s work]. . . is a remarkable achievement; however the latter part of his paper is so impenetrable that it is destined to be admired rather than read.

At the same time, the basic idea that underlies Enflo’s construction is a natural one in a sense that we will be better able to describe in Section 6.1. He constructs the space X as he goes, by putting a norm on the space of polynomials so that, with the resulting space completed to a Banach space, the shift operator (multiplication by the independent variable) has no invariant subspace. So here the space is complicated but the operator is simple. B. Beauzamy has published what is essentially an exposition of Enflo’s example, with some considerable simplifications. Counterexamples have also been given by C. Read; one of these is an example with a simple space (1 ) but a complicated operator. Related to the question of existence of invariant subspaces is the problem of determining all of the invariant subspaces of a given operator. Generally speaking, this is a very difficult problem, although there have been some notable successes. For example, the invariant subspaces of the shift operator Mz of multiplication by z on the Hardy space H 2 have been determined and have a beautiful structure (see [40]). The operator of multiplication by z on the Bergman space La2 (D) is known to have an extremely complicated lattice of invariant subspaces, and an understanding of this structure is matter of on-going work.

4.5 Introduction to the Spectrum We start with a definition, which will be of fundamental importance to us. Definition 4.26. If T : X → X is a bounded linear operator on a Banach space X, the set of complex numbers λ for which T − λ I is not invertible is called the spectrum of T . We will denote the spectrum of T by σ (T ). The spectrum of an operator on a Hilbert or Banach spaces contains vital information about the operator. It is an “invariant” of the operator in the sense of the following result. Proposition 4.27. If T is an operator in B(X) for a Banach space X, and if S is an invertible operator in B(X), then σ (T ) = σ (S−1 T S). Proof. If T − λ I is invertible, with inverse V , then S−1 T S − λ I = S−1 (T − λ I)S has inverse S−1V S. Conversely, if S−1 (T − λ I)S is invertible, then applying the first part we see that S[S−1 (T − λ I)S]S−1 = T − λ I is invertible. Since S−1 (T − λ I)S = S−1 T S − λ I, this completes the proof. When T1 and T2 in B(X) are related by T2 = S−1 T1 S for some invertible S ∈ B(X), we say that T1 and T2 are similar. Thus the last result says that similar operators have the same spectrum. The reader can show, by means of 2 × 2 matrices (that is, by operators in B(C2 )), that the converse is not true and two operators can have the same spectrum but fail to be similar.

4.5 Introduction to the Spectrum

97

It is helpful to think about how a complex number λ could get into σ (T ). Recall that an operator T − λ I is invertible if and only if T − λ I is bijective. So one way for λ to be a point of σ (T ) is for T − λ I to fail to be one-to-one. By linearity, this happens if and only if there is a vector g = 0 with T g = λ g, and thus λ is an eigenvalue of T . If λ ∈ σ (T ) but λ is not an eigenvalue of T , then it must be the case that T − λ I is not surjective. It sometimes helps to distinguish two ways this could happen: Either the range of T − λ I, while not all of X, is at least dense in X, or the closure of the range of T − λ I is a proper subspace of X. These two “parts” of the spectrum are called, respectively, the continuous spectrum and the residual spectrum. In some sense this last piece is the most intractable part of the spectrum, and we’ll see later that for certain classes of operators (for example, self-adjoint operators on a Hilbert space) the residual spectrum is empty. There are other useful ways of distinguishing various pieces of the spectrum; some of these (approximate eigenvalues, compression spectrum, essential spectrum) will be discussed in Sections 5.2 and 5.3. In all of this discussion the reader should keep in mind the much simpler situation for a linear operator on a finite-dimensional space, where the operator is bijective if and only if it is injective. In other words, for an operator on a finite-dimensional space, the spectrum is just the set of eigenvalues of the operator. Example 4.28. Consider the operator Mx of multiplication by ϕ (x) = x on the Hilbert space (L2 [0, 1], dx). In Exercise 4.15 you are asked to show that Mx has no eigenvalues. We claim, however, that each 0 ≤ λ ≤ 1 is in σ (Mx ). To see this, it is helpful to recall that an invertible operator is bounded below (meaning Ag ≥ δ g|| for some positive δ and all g; see Definition 2.24), and to observe that Mx − λ I = Mx−λ , the operator of multiplication by x − λ . If 0 < λ < 1 choose N sufficiently large that if n ≥ N then 1 1 En ≡ [λ − , λ + ] ⊆ [0, 1] n n and set  n gn = χE 2 n for all n ≥ N. The gn are unit vectors in L2 [0, 1] and (Mx − λ I)gn 2 = (x − λ )gn 2 =

n 2

 En

|x − λ |2 dx ≤

n 2

 2 2 1 1 = . n n n2

This computation shows that Mx − λ I is not bounded below, and hence not invertible. A similar argument, with En = [0, 1/n] or En = [1 − (1/n), 1] applies to show that λ = 0 and λ = 1 are also in σ (Mx ). Finally, we claim that no point outside of the interval [0, 1] can lie in σ (Mx ). If λ ∈ C\[0, 1], then

98

4 Compact Operators

1 ∈ L∞ [0, 1] x−λ and M1/(x−λ ) is a bounded operator which is inverse to Mx − λ I. Many further examples of concrete operators and their spectra will appear in later sections, when we have a bit more machinery at our disposal. The terminology “spectrum” comes from David Hilbert, who made major contributions to functional analysis initially motivated by a study of integral equations. Especially important were a collection of six papers written by Hilbert in the period 1904–1910 (and published together as a book in 1912 [19]). The fourth of this series of papers marks the beginning of the modern spectral theory. Here, in a general discussion of bilinear and quadratic forms, he generalized the concept of eigenvalue to that of the “spectrum”2 and began the study of the relationships between the operator T and its spectrum. Retrospectively, and in modern language, we can say that he studied self-adjoint operators on 2 , and an important aspect of this work was the discovery of ways to deal with the complications that arise when the continuous spectrum is not empty (and thus we are “outside” of the finite-dimensional case). This led to a description of any bounded self-adjoint operator on 2 which is a generalization of what we have done for compact self-adjoint operators in Section 4.3, and which leads to the spectral theorem as we shall discuss it in Chapter 6. When later it was discovered that the mathematical setting of self-adjoint operators on Hilbert space was a useful mathematical tool for theoretical physicists who were developing the then new theory of quantum mechanics, the spectra of these operators became related to the explanation of the “spectra” of atoms. Hilbert comments on this remarkable coincidence of terminology [38]: I developed my theory of infinitely many variables from purely mathematical interests, and even called it “spectral analysis” without any presentiment that it would later find an application to the actual spectrum of physics (p. 183).

Indeed, it is remarkable how the development of the theory of operators on Hilbert spaces occurred just as it was needed for the development of quantum mechanics. As A.M. Vershik writes in an essay on functional analysis in the twentieth century [45]: One might even conjecture that if the functional analysis of Hilbert spaces had not yet existed at the time when quantum mechanics arose, it would have been created out of necessity. For that reason, it is no exaggeration to say that the extremely close connection between the latest physics of the first half of the twentieth century and functional analysis gave the latter even greater authority (p. 441).

Hilbert defined the spectrum of T as the set of λ for which I − λ T is not invertible, which gives the reciprocals of what is now the commonly used definition.

2

4.6 The Fredholm Alternative

99

4.6 The Fredholm Alternative Our main goal in this section is to get information on the spectrum of a compact operator on a Hilbert space. Theorem 4.29. Suppose X is a Banach space and T ∈ B(X) is compact. If λ = 0, then T − λ I has closed range. Proof. For λ = 0, T − λ I = λ ( λ1 T − I). Since λ1 T is compact if T is, it suffices to prove the theorem for λ = 1. Suppose, for a contradiction, that the range of T − I is not closed. Define a map S from the quotient space X/ker (T − I) into X by S(x + ker (T − I)) = (T − I)x. By Exercise 3.26 in Chapter 3, we know that S is a well-defined, bounded linear map, which is one-to-one. The range of S is equal to the range of T − I, and hence by our assumption, the range of S is not closed. In Exercise 2.12 of Chapter 2 it was shown that a Hilbert space operator that is bounded below must have closed range, and it is easy to see that this same result holds in Banach spaces as well. Thus S is not bounded below, and so there must be (quotient space) unit vectors xn + ker (T − I) in X/ker (T − I) with S(xn + ker (T − I)) = (T − I)xn  → 0. By the definition of the coset norm, if xn + ker (T − I) = 1, then for any positive ε there exist yn ∈ ker (T − I) such that xn − yn  ≤ 1 + ε . Since xn − yn + ker (T − I) = 1, there is no loss of generality in assuming, say, that xn  ≤ 2 for all n. Compactness of the operator T then guarantees that T xn has a convergent subsequence, and hence we may assume (not relabeling this subsequence) that T xn → y for some y in X. Since (T − I)xn → 0, we must have xn → y, and by continuity, T xn → Ty. Thus Ty = y and y is in ker (T − I). Now we have a contradiction: Writing [xn ] for the coset xn + ker (T − I), we have [xn ] = 1 and [y] = 0 but also xn → y. The restriction λ = 0 in Theorem 4.29 is crucial; see Exercise 4.26. Theorem 4.30. Suppose that T is a compact operator on a Hilbert space H and let M j be the range of the operator (T − I) j for each j = 1, 2, . . .. There exists a positive integer j such that M j = M j+1 . Proof. By the previous theorem, M1 is closed. For j > 1, we may expand (I − T ) j by the binomial theorem to write (I − T ) j = I − jT +

j( j − 1) 2 T + · · · + (−1) j T j , 2

where A ≡ jT − j( j −1)/2T 2 +· · ·−(−1) j T j is compact, by Proposition 4.9. Apply the previous theorem to I − A to conclude that M j is closed for each j. Clearly M j+1 ⊆ M j ; suppose this containment is proper for each j. The quotients M j /M j+1 would each then have dimension at least 1, and for each j we can choose x j in M j with x j + M j+1  = 1. As in the proof of the preceding theorem, there is no

100

4 Compact Operators

loss of generality in assuming that x j  ≤ 2 for each j. We claim that T x j −T xk  ≥ 1 for j = k, contradicting the hypothesis that T is compact. To this end, suppose j < k so that j < j + 1 ≤ k < k + 1. We have • xk ∈ Mk ⊆ M j+1 , • (T − I)x j ∈ M j+1 by definition of M j+1 , and • (T − I)xk ∈ Mk+1 ⊆ M j+1 . Defining y ≡ (T − I)x j − (T − I)xk − xk , we see that y is in M j+1 , and the definition of the coset norm guarantees that x j + y ≥ 1. But x j + y = T x j − T xk , and we have verified our claim. The result of Theorem 4.30 is sometimes phrased as “T − I has finite descent if T is compact.” The next result is the main result of this section. It says that the nonzero points in the spectrum of a compact operator are always eigenvalues of the operator. Theorem 4.31. Suppose T is a compact operator on a Hilbert space H and λ = 0. If T − λ I is not invertible, then λ is an eigenvalue of T . Proof. Since T − λ I = λ ( λ1 T − I), there is no loss of generality in taking λ = 1. Thus we are given that T − I is not invertible. Suppose that 1 is not an eigenvalue of T . This means ker (T − I) = {0}, and T − I is one-to-one. Since it is not invertible, it must therefore fail to map onto H : (T − I)H is properly contained in H . Since T − I is one-to-one, it follows that (T − I)2 H is properly contained in (T − I)H , for if x0 fails to be in the range of T − I, then (T − I)x0 fails to be in the range of (T − I)2 . Continuing, we see that for each j, the range of (T − I) j+1 is properly contained in the range of (T − I) j . This is in contradiction to the conclusion of Theorem 4.30, and we are done. The next result, called the “Fredholm alternative,” summarizes what we have learned in this section. Notice how it captures results you know from linear algebra about linear maps on Cn . Theorem 4.32. Let T be a compact operator on a Hilbert space H . Suppose λ is a nonzero complex number. (a) (b)

If T − λ I is one-to-one, then T − λ I is invertible. If T − λ I maps H onto H , then T − λ I is invertible.

Proof. The first statement is Theorem 4.31. For the second, take adjoints. If T − λ I is onto, then T ∗ − λ I is one-to-one (by Exercise 2.16 in Chapter 2). From the first part of the theorem, T ∗ − λ I is invertible; the adjoint of its inverse provides the inverse to T − λ I. There is a pithy way of describing the conclusions of Fredholm alternative. Think of “(T − λ I)x = y” as an equation with “y” given and “x” as the unknown. The second conclusion in the Fredholm alternative says “if a solution exists for all y, then it is unique” while the first conclusion says “if the solution is unique, it exists.” Theorem 4.32 can be extended to the following result.

4.7 Exercises

101

Theorem 4.33. If T is compact in B(H ) and λ is a nonzero value, then dim ker (T − λ I) = dim [ran (T − λ I)]⊥ . Theorem 4.32 is the case where the common value of the two numbers is zero. We do not give the proof of Theorem 4.33 here, but refer the interested reader to Lemma 3.2.8 in [2]. Every result in this section has an exact Banach space analogue, and only minor modifications need to be made to obtain the proofs in this more general setting. In particular Theorems 4.30 and 4.31 go through exactly as before. For Theorem 4.32, one need only pay attention to the fact that the adjoint is defined slightly differently in the Banach space context, check that it is still true that ker A∗ = (ran A)⊥ , where now M ⊥ denotes the bounded linear functionals which are zero at each point of M, and recall that (T − λ I)∗ = T ∗ − λ I.

4.7 Exercises 4.1. Suppose that  · α and  · β are two equivalent norms on a vector space X. Show that if (X,  · α ) is complete, then so is (X,  · β ). 4.2. Suppose that X is an n-dimensional normed linear space over C. Show that there is a linear bijection T : X → Cn such that T and T −1 are continuous (in your choice of a norm for Cn ); in short, every n-dimensional normed linear space over C is isomorphic to Cn . 4.3. Suppose that X is a normed linear space, endowed with the metric topology, and suppose X contains a nonempty open set V such that V is compact. The goal of this problem is to show that this forces X to be finite-dimensional. (a) Without loss of generality we may assume that 0 ∈ V . Show that as x ranges over the set V , the open sets x + 12 V ≡ {x + 12 v : v ∈ V } form an open cover of V . By compactness, extract a finite subcover N  1 . xk + V 2 k=1 Define Y to be the span of the points x1 , x2 , . . . , xN . (b) Show that V ⊆ Y + 21j V for each positive integer j, and hence V⊆

∞ 

(Y +

j=1

(c) Show that

∞

1 (Y

+ 21j V ) = Y .

1 V ). 2j

102

4 Compact Operators

(d) From (b) and (c) and the fact that for any x ∈ X, a sufficiently small, but nonzero multiple of x will lie in V , conclude that X = Y , and thus that X is finitedimensional. 4.4. Suppose that M1 is a closed subspace and M2 is a finite-dimensional subspace in a normed linear space X. (a) Show that M1 + M2 ≡ {m1 + m2 : m1 ∈ M1 , m2 ∈ M2 } is a closed subspace of X. Hint: Argue that it is enough to consider M2 to be one-dimensional, M2 = {α x0 : α ∈ C}. Suppose mn + αn x0 → y where mn ∈ M1 , αn ∈ C. Show that {αn } is a bounded sequence of complex numbers, and extract a convergent subsequence αnk . Write mnk = (αnk x0 + mnk ) − αnk x0 . (b) Use (a) to give an alternate proof of the statement in Proposition 4.3 that a finitedimensional subspace in a normed linear space is closed. 4.5. This problem provides an alternate proof to Proposition 4.4. Suppose that T : X → Y is linear, where X and Y are normed linear spaces and X is finite-dimensional. Define  · β on X by xβ = max(xX , T xY ). (a) Check that  · β is a norm on X. (b) Argue that T : (X,  · β ) → Y is continuous, and hence that so is T : X → Y in the original norm on X. 4.6. Let X be a Banach space and suppose T1 , T2 , S are bounded linear operators from X into X, with T1 and T2 compact. Show that T1 + T2 , α T1 , ST1 , and T1 S are all compact (α any scalar). If F is a finite rank operator, show that SF and FS are finite rank as well. 4.7. Find the error in the following “proof” that the compact operators on a Banach space X are closed in the bounded operators on X. Alleged proof: Suppose Tn is compact for each n and suppose further that Tn − T  → 0 for some bounded linear operator T . To show that T is compact, we want to show that for an arbitrary bounded sequence {xn } in X, {T xn } has a convergent subsequence. Fix such a sequence {xn } and let M be a bound for it: xn  ≤ M for all n. Now choose an ε > 0, and find K sufficiently large that TK − T  ≤

ε . 3M

We are given that the operator TK is compact, so we can find a subsequence {xn j } of our sequence {xn } so that TK (xn j ) converges. But a convergent sequence must be a Cauchy sequence, so if n j and nk are sufficiently large, say if n j , nk ≥ N, then

ε TK (xn j ) − TK (xnk ) < . 3 We claim that T (xn j ) converges. Since we are in a Banach space, to verify this, it is enough to show that {T (xn j )} is a Cauchy sequence. To this end, notice that for n j , nk ≥ N we have

4.7 Exercises

103

T (xn j ) − T (xnk ) ≤ T (xn j ) − TK (xn j ) + TK (xn j ) − TK (xnk ) + TK (xnk ) − T (xnk ) ε ≤ T − TK  · M + + TK − T  · M 3 ≤ ε. This shows that {T (xn j )} is Cauchy, and hence it converges, as desired. 4.8. If A ∈ B(H ) is a diagonal operator with diagonal {αn }, show that if A is compact, then limn→∞ αn = 0. 4.9. This problem builds on Exercise 2.6 in Chapter 2. Show that every finite rank operator T in a Hilbert space H can be described as n

Th =

∑ h, x j y j

j=1

for orthonormal vectors x1 , x2 , . . . , xn and vectors y1 , y2 , . . . , yn . 4.10.(a) Give an example of a compact operator which is not Hilbert–Schmidt. (Hint: look for a diagonal operator with this property.) (b) Show that no compact operator on an infinite-dimensional Hilbert space is invertible. (Exercise 4.24 below extends this result to Banach spaces). (c) Show that if T is compact in B(H ), where H is a Hilbert space, and λ = 0 is an eigenvalue of T , then ker (T − λ I) is finite-dimensional. 4.11. For λ ∈ C we abbreviate T − λ I (I the identity operator) by T − λ . (a) Suppose that T is a normal operator in B(H ). Show that (T − λ )h = (T − λ )∗ h for all h in H and all scalars λ . Hence ker(T − λ ) = ker(T − λ )∗ . (b) Show that if T is normal, then eigenvectors corresponding to distinct eigenvalues are orthogonal. (c) State and prove a version of Theorem 4.21 for compact normal operators. 4.12. If T ∈ B(H ) for some (complex) Hilbert space H and T h, h is real for all h ∈ H , show that T is self-adjoint. 4.13. In the notation of Corollary 4.25, show that for a given h ∈ H , we can solve the equation T f = h for f if and only if h ⊥ ker T and 1

∑ λn2 | h, en |2 < ∞. Find all such solutions f under these assumptions.

104

4 Compact Operators

4.14. Consider the weighted shift operator on 2 given by 1 1 W (x1 , x2 , x3 , . . .) = (0, x1 , x2 , x3 , . . .). 2 3 Show that W is compact, but W has no eigenvalues. Find a nontrivial closed invariant subspace for W . 4.15. Let Mx be the multiplication operator acting on L2 ([0, 1], dx) by Mx ( f ) = x f . Note that Mx is self-adjoint. Show that it has no eigenvalues, but many reducing subspaces. 4.16. Show that the Volterra operator V of indefinite integration on L2 ([0, 1], dx), defined by  x f (t)dt, V f (x) = 0

is compact. 4.17. Show that the subspaces Mα = { f ∈ L2 [0, 1] : f = 0 almost everywhere on [0, α ]} for any 0 ≤ α ≤ 1 are invariant subspaces for the Volterra operator V (defined in Exercise 4.16). In fact, every invariant subspace of V is of the form Mα for some α . This is a deep result; a proof can be found in [35]. 4.18. Suppose that M is a closed subspace of H so that H = M ⊕ M ⊥ . If A is in B(H ), we can write A as a matrix with operator entries   X Y A= , ZW where X ∈ B(M), Y ∈ B(M ⊥ , M), Z ∈ B(M, M ⊥ ), and W ∈ B(M ⊥ ). If M is an invariant subspace for A, what does this tell you about Z? If M is reducing subspace for A, what further information do you have about the operator entries of this matrix? 4.19. If {en } is an orthonormal sequence in a Hilbert space H , and T ∈ B(H ) is compact, show that Ten → 0. 4.20. Show that there is no nonzero multiplication operator on L2 (T, dx/(2π )) that is Hilbert–Schmidt. Is there a multiplication operator on L2 (T, dx/(2π )) that is compact? 4.21. Show that if T ∈ B(H ) is compact, and M is a closed invariant subspace of T , then the restriction of T to M is compact. 4.22. Show that a bounded linear operator T on a Hilbert space H which is selfadjoint and satisfies T 2 = T is an orthogonal projection onto its range.

4.7 Exercises

105

4.23. For an analytic function ϕ mapping the unit disk into itself, define the linear composition operator Cϕ by Cϕ ( f ) = f ◦ ϕ for f analytic on D. (a) Show that if Cϕ is Hilbert–Schmidt on the Bergman space La2 (D), then 

1 dA(z) < ∞. (1 − | ϕ (z)|2 )2 D

(b) Show that if Cϕ is bounded on La2 (D) and 

1 dA(z) < ∞, 2 2 D (1 − |ϕ (z)| )

then Cϕ is Hilbert–Schmidt. (c) Give an example of a compact composition operator on La2 (D). 4.24. Show that if X is an infinite-dimensional Banach space, then no bounded linear operator on X can be both compact and invertible. 4.25. Show that the result of Exercise 4.10(c) also holds for a compact operator on a Banach space. 4.26. Show that a compact operator A on a Banach space X can only have closed range if its range is finite-dimensional. 4.27. Suppose that A is a compact operator on a Banach space X. Show that if A2 = A, then the range of A is finite-dimensional. 4.28. Are the Hilbert–Schmidt operators a closed subspace of B(H )? 4.29. Suppose that T is a compact operator in B(H ) and λ = 0. Show that there exists a positive integer k so that ker (T − λ I)k = ker (T − λ I)k+1 . This is sometimes described as “T − λ I has finite ascent,” since for any operator A ∈ B(H ) we have ker A ⊆ ker A2 ⊆ ker A3 ⊆ · · ·.

Chapter 5

Banach and C∗ -Algebras

In 1943, a paper, written by I.M. Gelfand and M. Neumark, “On the imbedding of normed rings into the ring of operators in Hilbert space,” appeared (in English) in Mat. Sbornik. From the vantage point of a fifty year history, it is safe to say that the paper changed the face of modern analysis. R. Kadison ([25], p. 21).

In this chapter, our Banach spaces will be equipped with some additional structure which comes from a multiplication operation; that is, a Banach space A here will permit the multiplication of two vectors. This multiplication will be required to satisfy the following properties: (1) (2) (3) (4)

a(bc) = (ab)c (a + b)c = ac + bc a(b + c) = ab + ac λ (ab) = (λ a)b = a(λ b)

for all a, b, c in A and all scalars λ . Conspicuously absent from this list is any requirement of commutativity for this new multiplication operation, as well as the requirement that there be a multiplicative unit, i.e., a vector I such that aI = Ia = a for all a in A . We will impose these additional requirements (particularly the latter) from time to time, but at the moment neither is required. The terminology “complex algebra” is used for a vector space over C having properties (1)–(4) above; if a unit exists for the multiplication operation, we’ll say the algebra is “unital.” A Banach algebra is a complex algebra A with a norm making A into a Banach space and satisfying ab ≤ ab. Note this norm property guarantees that multiplication, as a map from A × A into A , is continuous: if an → a and bn → b then an bn → ab. This follows by writing an bn − ab = (an − a)bn + a(bn − b). When A is unital, we assume I = 1; see Exercise 5.2. The final layer of structure we will impose on some of the Banach algebras to be studied comes from the notion of an involution. An involution, on a Banach algebra A , is a map a → a∗ of A into A satisfying (1) (2) (3)

(a∗ )∗ = a (ab)∗ = b∗ a∗ (λ a + b)∗ = λ a∗ + b∗

B.D. MacCluer, Elementary Functional Analysis, DOI 10.1007/978-0-387-85529-5 5, c Springer Science+Business Media, LLC 2009 

107

5 Banach and C∗ -Algebras

108

for a, b ∈ A and λ a scalar. We call a∗ the adjoint of a. Finally, a C∗ -algebra is a Banach algebra with an involution such that a∗ a = a2 . We will call this the C∗ -identity. One way to motivate the “naturalness” of this last definition is to recall that for a bounded linear operator A on a Hilbert space H we have already observed A∗ A = A2 (Proposition 2.14). It is occasionally helpful to note that a Banach algebra with involution satisfying the inequality a∗ a ≥ a2 for all a ∈ A is a C∗ -algebra, meaning we get the inequality in the other direction for free. You are asked to provide the proof for this in Exercise 5.1.

5.1 First Examples Let us look at some examples, which show that these new definitions are all quite natural. Example 5.1. Consider C with the usual multiplication, absolute value as norm, and conjugation as involution: z∗ = z. The C∗ -identity is the familiar statement |zz| = |z|2 , and C is a commutative C∗ -algebra with unit, 1. Example 5.2. Let X be any compact Hausdorff space, and consider the Banach space C(X) of all continuous, complex-valued functions on X in the supremum norm, with pointwise-defined multiplication. This is a commutative Banach algebra, with the constant function 1 serving as the multiplicative unit. Defining an involution on C(X) by f ∗ (x) = f (x) makes C(X) into a C∗ -algebra. We’ll see later that every commutative unital C∗ -algebra is “isometrically isomorphic” to C(X) for some choice of a compact Hausdorff space X. Example 5.3. Now let X = R and consider the Banach space C0 (R) of continuous complex-valued functions that vanish at ∞ (meaning limx→±∞ f (x) = 0, or equivalently, that {x : | f (x)| ≥ ε } is compact for every ε > 0) in the supremum norm. Define multiplication pointwise and involution just as in the previous example. Then C0 (R) is a commutative, but nonunital, C∗ -algebra. This example can be generalized by replacing the real line by any noncompact but locally compact Hausdorff space X; analogously to the last comment in the previous example, every commutative nonunital C∗ -algebra is C(X) for some locally compact Hausdorff space X. Example 5.4. Starting with a σ -finite measure space (X, M, µ ), the Banach space L∞ (X, µ ), with multiplication and involution defined as in the last two examples, is a commutative, unital C∗ -algebra. (Strictly speaking, to define the multiplication of two elements of L∞ (X, µ ) we choose a representative of each and define its pointwise product to be a representative of the product element.) Note it would not do to replace “∞” by “p < ∞” in this example, as the multiplication of two L p functions need not be in L p for finite p.

5.1 First Examples

109

Example 5.5. Our most important example is the Banach algebra of all bounded linear operators on a Hilbert space H , normed by the operator norm A = sup{Ah : h = 1}, and with multiplication defined by composition (AB)(h) = A(B(h)). This is a noncommutative (when the dimension of H is at least two) Banach algebra with identity I. Defining A∗ to be the usual operator adjoint provides an involution on B(H ) under which we have a C∗ -algebra, as noted in Chapter 2. In the special case that H = Cn , then B(H ) is identified with the n × n matrices, and we will often denote this by Mn . When H is replaced by a Banach space X, B(X) is a Banach algebra. Even though many of the classical Banach spaces are in fact Banach algebras under a natural multiplication, the conscious exploitation of this fact was rather long in coming. Riesz, writing in 1913, looked explicitly at the product of operators on a Hilbert space, and was at least implicitly aware of the inequality AB ≤ AB. By 1930, the concept of “rings of operators” came under explicit study, and beginning in 1936 an important series of papers by Francis Murray and John von Neumann, titled “On Rings of Operators,” developed the theory of what are now called von Neumann algebras. These are certain kinds of C∗ -subalgebras of B(H ), and a particular motivation for their study was to provide the “right” mathematical framework for the study of “observables” in quantum mechanics. Von Neumann was a brilliant and prolific mathematician who made fundamental contributions to many areas of both pure and applied mathematics. He (along with Albert Einstein and Kurt G¨odel) was part of the first faculty at the Institute for Advanced Study in Princeton. Peter Lax, in the forward to a recently published collection of letters written by von Neumann, says ...had he lived a normal span of years1 , he would certainly have been a recipient of a Nobel Prize in economics. And if there were Nobel Prizes in computer science and mathematics, he would have been honored by these, too. So the writer of these letters should be thought of as a triple Nobel laureate, or possibly, a 3 12 -fold winner, for his work in physics, in particular quantum mechanics ([37], p. xiii).

His work with Murray is among his most influential, at least on the pure mathematics side. Curiously, this work predates much of the foundational work on Banach algebras that we will look at in the next sections (much of which is due to I. Gelfand). Von Neumann showed a prodigious talent as a young child for calculation and solving problems. According to a biographical article on von Neumann written by Halmos [16], At the age of 6 he could divide two eight digit numbers in his head; by 8 he had mastered the calculus; by 12 he had read and understood Borel’s Th´eorie des Fonctions (p. 383).

Stories about his astonishing calculational abilities recur throughout his life. His biographer, N. Macrae, tells the following anecdote [30]:

1

Von Neumann died in 1957, at the age of 53, of cancer.

110

5 Banach and C∗ -Algebras

When calculating a problem while sitting, he was apt to stare at the ceiling muttering, with an almost frighteningly blank face. He did this when the Rand Corporation asked whether his computers could be modified to tackle a particular problem, which—as Rand staff explained to him for two hours on blackboards and with graphs—would understandably be beyond computers in their present state. For two or three minutes, Johnny “stared so blankly that a Rand scientist later said he looked as if his mind had slipped his face out of gear. Then he said ‘Gentlemen, you do not need the computer, I have the answer”’(p. 9).

Any norm-closed subalgebra of a C∗ -algebra which is also closed under adjoints is again a C∗ -algebra. An important example is the subalgebra of compact operators K (H ) in B(H ) (Theorem 4.10 and Proposition 4.12 are relevant here); it is noncommutative and has no unit when the dimension of H is infinite. The term B∗ -algebra also appears in the literature, for what we call a C∗ -algebra. At one point in time, the terminology C∗ -algebra was reserved for a closed subalgebra of B(H ) which was also closed under the ∗ operation. The Gelfand–Naimark theorem established the fact that every B∗ -algebra was “isometrically ∗ - isomorphic” to a closed ∗ -subalgebra of B(H ) for some choice of a Hilbert space H , and the need for separate terminology disappeared, with “C∗ -algebra” winning out. Since B(H ) is our most important example of a C∗ -algebra, and in view of the Gelfand–Naimark theorem just discussed, we will henceforth use uppercase letters A, B,C, . . . to denote elements of a generic Banach or C∗ -algebra, and write I for the multiplicative unit, if there is one. We will also carry over some familiar terminology from the B(H ) setting and call an element A of a C∗ -algebra self-adjoint if A = A∗ and normal if AA∗ = A∗ A. Thus, for example, the self-adjoint elements of C(X) as in Example 5.2 above are the real-valued functions in C(X).

5.2 Results on Spectra For a bit, we don’t need the C∗ -structure, so in this section we work in Banach algebras. Since we’ll be concerned with invertibility, we will assume our Banach algebras to be unital. As noted above, the bounded linear operators B(X), where X is a Banach space, form a Banach algebra, under the usual multiplication of operators (that is, composition). Definition 5.6. Suppose A is a Banach algebra with unit I. We say that A ∈ A is invertible if there exists B in A with AB = BA = I. We could talk separately about “left” and “right” inverses for an element in a Banach algebra. It is easy (and useful!) to note that if A has both a right inverse B and left inverse C, then B = C, since C = CI = C(AB) = (CA)B = IB = B. Thus when convenient, we can show that an element is invertible by separately exhibiting a left and right inverse. For n × n matrices, the existence of either a right or left inverse guarantees invertibility of the matrix. The analogous statement does not hold more generally, though. For example, the forward shift S on 2 has left inverse S∗ but is not invertible.

5.2 Results on Spectra

111

In the next result, the topology in use in A is of course the metric topology which comes from the norm. Theorem 5.7. Suppose A is a unital Banach algebra and let G denote the invertible elements of A . Then G is an open set in A . To prove this theorem, we begin with a lemma which says that the open ball of radius 1 about the identity is contained in G . Lemma 5.8. If B ∈ A and I − B < 1, then B is invertible, and its inverse is given k by ∑∞ k=0 (I − B) . Notice that the formula for B−1 follows formally from a geometric series type k 0 manipulation: B−1 = [I − (I − B)]−1 = ∑∞ k=0 (I − B) , where we interpret (I − B) to be I. Proof (Lemma 5.8). Let C = I − B so that C = r < 1 and Cn  ≤ Cn = rn . Since ∞ n n r < 1 we have ∑∞ 0 C  < ∞. This says the partial sums of ∑0 C form a Cauchy sequence, and hence by completeness they converge in A . Denote ∑nk=0 Ck by Zn k 2 n n+1 → I. and ∑∞ k=0 C by Z. Since Zn = I +C +C + · · ·C , we have Zn (I −C) = I −C On the other hand, Zn (I −C) → Z(I −C), so we must have Z(I −C) = I. Similarly, (I −C)Z = I, and I −C is invertible, with inverse Z. Since I −C = I − (I − B) = B we are done. With the lemma in hand, we can make short work of the proof of Theorem 5.7. Proof (Theorem 5.7). Suppose A0 ∈ G with inverse B0 . We claim that for all A satisfying A − A0  < B0 −1 , A is invertible. To see this, note that A − A0  < B0 −1 implies that B0 A−B0 A0  < 1; that is, I −B0 A < 1. Invoking Lemma 5.8, we find that B0 A is invertible, say (B0 A)−1 = C. Since I = C(B0 A) = (CB0 )A, we see that A is left invertible. Similarly, A − A0  < B0 −1 implies 1 > AB0 − A0 B0  = AB0 − I and AB0 is invertible, with inverse, say, D. Thus A has right inverse B0 D. As we have already observed, the existence of a left inverse and a right inverse for A shows that A ∈ G , and the proof is complete. The map from G to G sending A to A−1 is continuous; see Exercise 5.3. We have looked at the notion of the spectrum of an operator in Section 4.4; it is straightforward to formulate the definition more generally, for an element of a unital Banach algebra. Definition 5.9. Let A ∈ A , where A is a unital Banach algebra. The spectrum of A, denoted σ (A), is {λ ∈ C : A − λ I is not invertible in A }. We often write A − λ for A − λ I. Of course, when A is B(X) for some Banach space X, the above definition is precisely our earlier notion of the spectrum of a bounded linear operator. As another example, suppose that A is the Banach algebra is C(X) for some compact Hausdorff space X. The spectrum of f in C(X) is the set of complex numbers λ for which f (x) − λ = 0 for some x ∈ X, that is to say, σ ( f ) = f (X), the range of f .

5 Banach and C∗ -Algebras

112

The next result collects some of the basic properties of the spectrum of an element in a unital Banach algebra, the deepest of which is the statement that the spectrum is always nonempty. A moment’s reflection on the linear algebra roots of this statement puts this in some perspective. When A is an n × n complex matrix (i.e., an element of the Banach algebra Mn ), the spectrum of A is exactly the set of eigenvalues of A. The statement that the spectrum is nonempty becomes, in this setting, the statement that every n × n matrix has a (complex) eigenvalue, a nontrivial fact whose usual proof makes use of the fundamental theorem of algebra. Theorem 5.10. Suppose A is a unital Banach algebra and let A ∈ A . The spectrum of A is a nonempty, compact subset of C, which is contained in the closed disk {λ : |λ | ≤ A}. Moreover, the map which sends z to (z − A)−1 is an A -valued strongly analytic function on the open set C\σ (A). The last part of the statement of the theorem requires some explanation. How do we define analyticity for vector-valued functions? For an open set Ω in the complex plane and a Banach space A define the derivative of the vector-valued function f : Ω → A at z0 ∈ Ω to be f  (z0 ) = lim

h→0

f (z0 + h) − f (z0 ) h

if it exists in A . The quotients ( f (z0 + h) − f (z0 ))/h are vectors in A , and the limit is taken in the norm topology of A , as h tends to 0 in C. We say f is strongly analytic in Ω if f  exists and is continuous in Ω . Another natural way to contemplate defining analyticity for a Banach space-valued function is as follows: Say that f : Ω → A is weakly analytic if ϕ ◦ f is analytic in the ordinary sense for every bounded linear functional ϕ ∈ A ∗ . It is easy to show that a Banach space-valued strongly analytic function is weakly analytic (Exercise 5.4); remarkably the converse is also true, so that these two definitions of analyticity actually coincide. Since we won’t need this latter fact, we omit the proof here. To prove Theorem 5.10 will need a “Liouville-type theorem” for vector-valued analytic functions, which we turn to next. Theorem 5.11. If f : C → A is weakly analytic, where A is a Banach space, and bounded in C, then f is constant. Proof. We are given that  f (z) ≤ M < ∞ for all z ∈ C. If ϕ is arbitrary in A ∗ , then ϕ ◦ f is entire and |ϕ ( f (z))| ≤ ϕ  f (z) ≤ ϕ M for all z ∈ C. Thus for every bounded linear functional ϕ on A , ϕ ◦ f is a bounded entire function in the complex plane, and hence constant. We claim this implies that f is constant. Suppose not, and find z1 , z2 with f (z1 ) = f (z2 ). By Corollary 3.4 to the Hahn–Banach theorem, A ∗ separates the points of A and we may find ϕ ∈ A ∗ with ϕ ( f (z1 )) = ϕ ( f (z2 )), a contradiction. This verifies the claim and completes the proof.

5.2 Results on Spectra

113

We’re now ready to prove Theorem 5.10. Proof (Theorem 5.10). Let |λ | > A and write A − λ = λ ( λA − I). By Lemma 5.8 we can see that I − A/λ , and hence A − λ , is invertible. Thus the spectrum of A is contained in the closed disk {λ : |λ | ≤ A}. Next we show that σ (A) is closed by showing its complement is open. If A − λ is invertible, we want to show that for some ε > 0, A − µ is invertible provided |λ − µ | < ε . Since the set G of invertible elements of A is open, we can find an ε > 0 so that if B − (A − λ ) < ε , then B is invertible. For this ε , |λ − µ | < ε implies A − µ is invertible. This shows that the complement of σ (A) is open. Being closed and bounded in C, σ (A) is compact. Define F : C\σ (A) → A by F(λ ) = (λ − A)−1 . We claim that F is (strongly) analytic in C\σ (A). For h in C sufficiently small so that λ + h stays in the open set C\σ (A) we have (λ + h − A)−1 − (λ − A)−1 = (λ + h − A)−1 [(λ − A) − (λ + h − A)](λ − A)−1 so that

F(λ + h) − F(λ ) = −(λ + h − A)−1 (λ − A)−1 . h Continuity of the inverse (Exercise 5.3) shows that lim

h→0

F(λ + h) − F(λ ) = −[(λ − A)−1 ]2 h

verifying the analyticity of F. Finally we turn to the assertion that the spectrum of A is nonempty. If it were empty, then the A -valued function F as just defined is analytic in all of C. Moreover, for |λ | > A, F(λ ) = (λ − A)−1 = λ −1 (I − A/λ )−1 , which tends to 0 as |λ | → ∞. / then F is a bounded, entire A -valued function, and hence constant Thus if σ (A) = 0, by Theorem 5.11, which is clearly a contradiction. The next result, called the Gelfand–Mazur theorem, is due independently to Gelfand (1941) and Mazur (1938). It uses Theorem 5.10 to show that the only unital Banach algebra which is also a division algebra is C. An isomorphism between Banach algebras is a bijective linear map which is also multiplicative; it is isometric if it also preserves norms. Theorem 5.12. If A is a unital Banach algebra in which each nonzero element is invertible, then A is isometrically isomorphic to C. Proof. Let A ∈ A and suppose λ1 , λ2 are two distinct complex numbers. At least one of A − λ1 , A − λ2 is invertible (since both can’t be 0). On the other hand, σ (A) is nonempty, so σ (A) consists of exactly one complex number for each A ∈ A ; call it λ (A). Now A − λ (A)I = 0 or A = λ (A)I. The mapping sending A to λ (A) is an isometric isomorphism of A onto C; verification of the routine details of this statement is left to the reader.

5 Banach and C∗ -Algebras

114

For a polynomial p(z) = cn zn + · · ·+ c1 z + c0 and an element A of a unital Banach algebra, we write p(A) for cn An +· · ·+c1 A+c0 I and p(σ (A)) for the set{p(λ ) : λ ∈ σ (A)}. Lemma 5.13. The product Π1n (A− λ j I) is invertible if and only if each of the factors A − λ j I is invertible. Proof. The “if” direction of the lemma is obvious. Now suppose Π1n (A − λ j I) is invertible, with inverse S. Fix a j, with 1 ≤ j ≤ n, and set P = Πm= j (A − λm I). Since A − λm I and A − λk I commute for every m and k, we have SP(A − λ j I) = I. This shows that A − λ j has a left inverse. Similarly, (A − λ j I)PS = I and A − λ j I has a right inverse. As we have observed, this guarantees that A − λ j I is invertible. A spectral mapping theorem is a statement of the form σ (p(A)) = p(σ (A)) for p in some class of functions and A an element of a unital Banach algebra. Our first spectral mapping theorem uses the class of polynomial functions. Theorem 5.14. Suppose A is a unital Banach algebra, A is an element of A , and p is a polynomial. We have

σ (p(A)) = p(σ (A)). Proof. The result is easy when p is a constant, so we assume p has degree at least one. We’ll show the two inclusions: p(σ (A)) ⊆ σ (p(A)) and the reverse. For the first, let λ ∈ σ (A) and factor p(z) − p(λ ) as c(z − λ )(z − λ2 ) · · · (z − λn ), where c = 0 and the λ j are complex numbers; note that we have used the fact that λ is a root of p(z) − p(λ ). Defining λ1 = λ , we then write p(A) − p(λ )I = c(A − λ1 I)(A − λ2 I) · · · (A − λn I). The fact that A − λ j I and A − λk I commute is being used here. By Lemma 5.13, invertibility of the product Π1n (A − λ j I) is equivalent to invertibility of each factor; since λ1 = λ ∈ σ (A), this says that p(A) − p(λ )I is not invertible and hence p(λ ) is in σ (p(A)). To see that σ (p(A)) ⊆ p(σ (A)), let µ ∈ σ (p(A)). Factor the polynomial p(z) − µ as c(z− β1 ) · · · (z− βn ) so that p(A)− µ I = c(A− β1 I) · · · (A− βn I). Now Lemma 5.13 says that some A − β j I is not invertible, so for some j, β j ∈ σ (A). But p(β j ) − µ = 0, so we have realized µ as p(β j ) where β j ∈ σ (A), and the proof is complete. As a simple example to illustrate this spectral mapping theorem, suppose that A is an element of a unital Banach algebra that satisfies A2 = A. This will imply that σ (A) ⊆ {0, 1}. To see that this follows directly from Theorem 5.14, let p(z) = z2 − z

5.2 Results on Spectra

115

so that p(A) = 0, and hence σ (p(A)) = {0}. We must then have p(λ ) = 0 for all λ ∈ σ (A), and thus σ (A) ⊆ {0, 1}. When A is a bounded operator on a Hilbert space H there are extensions of Lemma 5.13 which allow us to prove spectral mapping theorems for various “parts” of the spectrum of A. Recall that by Theorem 2.25, a bounded linear operator on a Hilbert space is invertible if and only if it is bounded below and has dense range. Thus a complex number λ is in the spectrum of A ∈ B(H ) if and only if A − λ I is not bounded below and/or A − λ I does not have dense range. This tells us that σ (A) is composed of two possibly overlapping sets, pictured below:

σap (A) ≡ {λ : A − λ I is not bounded below} (called the approximate point spectrum of A) and

Γ (A) ≡ {λ : A − λ I does not have dense range} (called the compression spectrum of A). The approximate point spectrum consists of two disjoint pieces, the eigenvalues of A (denoted σ p (A)) and the complement of σ p (A) in σap (A). This decomposition is schematically represented below. For further discussion of these parts of the spectrum of A, see Exercises 5.5 and 5.8.

σ p (A)

σap (A)

Γ (A)

FIGURE 5.1: Parts of the spectrum Exercise 5.7 outlines an argument to show that for every polynomial p, p(σap (A)) = σap (p(A)); this is a “spectral mapping theorem” for the approximate point spectrum. Exercise 5.9 outlines a proof of an “inversion spectral mapping theorem,” which says that for an invertible operator A ∈ B(H ),   1 −1 −1 σ (A ) = [σ (A)] ≡ : λ ∈ σ (A) . λ We know that the spectrum of any element A in a unital Banach algebra is compact, nonempty, and contained in the closed disk with radius A centered at the origin. We define the spectral radius of A to be

5 Banach and C∗ -Algebras

116

r(A) ≡ max{|λ | : λ ∈ σ (A)}. Note that r(A) = 0 is equivalent to σ (A) = {0}, but that r(A) = 0 does not necessarily imply A = 0; it is easy to find a counterexample in M2 . There is a remarkable formula, due to Gelfand, for r(A). This formula relates the algebraic property inherent in its definition (invertibility) to An , a metric property. Theorem 5.15 (Spectral Radius Formula). For A ∈ A , a unital Banach algebra, r(A) = limn→∞ An 1/n . The existence of the limit is part of the proof. We will take a look at some examples and applications before proceeding to the proof. Example 5.16. Suppose our Banach algebra is C(X) for some compact Hausdorff space X. We know σ ( f ) = range f = f (X), so r( f ) =  f ∞ . Thus the spectral radius 1/n formula clearly holds in this setting, since  f n ∞ = ( f ∞ )n and  f n ∞ =  f ∞ for all n. Example 5.17. If A is a self-adjoint element of a unital C∗ -algebra, then we have A2  = A∗ A = A2 and A2  = A2 . Since A2 is also self-adjoint, we may replace A by A2 to get A4  = A2 2 = A4 . Continuing, an induction argument will show n

n

A2  = A2 , and the spectral radius formula gives the conclusion that r(A) = A for any selfadjoint A. Much later we will see that the same conclusion is true more generally for normal elements. Suppose that A is a unital Banach algebra, and that B is closed subalgebra of A containing the unit I. Let B be an element of B. It is certainly possible for the spectrum of B, relative to B, to be different from the spectrum relative to A . (This issue is explored further in Exercise 5.17) However, by the spectral radius formula, the spectral radius must be the same whether we think of B as an element of A or B, since Bn  has nothing to do with which of the two algebras are being considered. In Section 5.6, we will see that if A is a C∗ -algebra with unit I, and B is a C∗ subalgebra of A containing I, then σA (B) = σB (B) for any B ∈ B; this is referred to as “spectral permanence” in C∗ -algebras. We are ready to give the proof of the spectral radius formula. Proof (Theorem 5.15). We begin with some observations that will clarify what we need to do. If λ is in σ (A), then by the spectral mapping theorem (Theorem 5.14)

5.2 Results on Spectra

117

λ n ∈ σ (An ) for any positive integer n. This tells us that |λ n | ≤ An  and thus we have |λ | ≤ An 1/n for all nonnegative integers n. It follows that r(A) ≤ inf An 1/n ≤ lim inf An 1/n . n

n→∞

Thus if we can show lim supn→∞ An 1/n ≤ r(A) we will have lim sup An 1/n ≤ r(A) ≤ lim inf An 1/n , n→∞

n→∞

forcing equality throughout and giving the desired result. To this end, let ∆ be the open disk in C centered at 0 and having radius 1/r(A) (which we interpret as ∞ if r(A) = 0). We make a few observations to get started: • The map z → (z − A)−1 is (strongly) analytic in C\σ (A), and has limit 0 as |z| → ∞, by Theorem 5.10 and its proof. • If λ is in ∆ \{0} then 1/λ is in C\σ (A), since |1/λ | > r(A). • The map λ → (I − λ A)−1 is strongly analytic in ∆ \{0} and continuous at 0, since (I − λ A)−1 = [λ ( λ1 I − A)]−1 . Thus for each bounded linear functional ϕ in A ∗ , the C-valued function f (λ ) ≡ ϕ [(I − λ A)−1 ] is analytic in ∆ \{0} and continuous at 0, hence analytic in ∆ . It must therefore have a power series representation in ∆ , which we determine next. If |λ |
0 such that ST h ≥ ch for all h ∈ (ker ST )⊥ . From this it follows that T h ≥

c h S

for all h ∈ (ker ST )⊥ . The reader can now easily show, using completeness, that {T h : h ∈ (ker ST )⊥ } is a closed subspace of H . Now write K for the compact operator ST − I. Then ker (ST ) = ker (I + K), where ker (I + K) is finite-dimensional by Exercise 4.10 in Chapter 4, and so T (ker ST ) is a finite-dimensional subspace of H . Thus we have a finite-dimensional subspace {T h : h ∈ ker (ST )}, and a closed subspace

{T h : h ∈ ker (ST )⊥ }.

We may apply Exercise 4.4 in Chapter 4 to conclude that {T h : h ∈ ker (ST )} + {T h : h ∈ ker (ST )⊥ } is a closed subspace. Since this is obviously the range of T , we are done.



We can rephrase Theorem 5.59 as “T ∈ B(H ) is Fredholm if and only if T + K is invertible in the Calkin algebra B(H )/K (H ).” As an immediate consequence we see that the product of two Fredholm operators is Fredholm. This result has a purely algebraic proof as well; see [41]. The terminology “Fredholm operator” recognizes the pioneering work of Erik Fredholm. In 1903 he published a paper that, in modern language, dealt with equations of the form  f (s) −

b

k(s,t) f (t)dt = g(s),

(5.3)

a

where k(s,t) is in L2 ([a, b] × [a, b]) and f , g are in L2 [a, b]. A natural question to ask is: For which g does a solution f exist to this equation, and when a solution exists for a particular g, can the solutions be described? From our modern perspective we can think of Equation (5.3) as (I − K) f = g,

5 Banach and C∗ -Algebras

146

where K is the integral operator with kernel k(s,t) as defined in Section 2.1. Furthermore, as we know from Theorem 4.16, the operator K is compact (in fact, Hilbert– Schmidt). The extent to which solutions to Equation (5.3) are not unique is measured by the dimension of the kernel of I − K, and the extent to which solutions (with g being given and f being the unknown) fail to exist is measured by the dimension of [ran (I − K)]⊥ . Notice how Theorem 4.32 elaborates on this: If the dimension of the kernel of I − K is zero, then I − K is invertible and for each g a unique solution f exists.

5.9 Exercises 5.1. Show that a Banach algebra A with an involution satisfying A∗ A ≥ A2 is a C∗ -algebra, meaning that equality holds in this inequality. 5.2. Suppose that A is a C∗ -algebra. (a) Show that if A has a unit, it is unique (call it I); furthermore I ∗ = I and I = 1 (provided A = 0 for some A ∈ A ). (b) Suppose A is unital. Show that if A is invertible, so is A∗ , with (A∗ )−1 = (A−1 )∗ . (c) Every A ∈ A can be written as A = X + iY where X and Y are self-adjoint. (d) If A is unital and U is unitary (meaning UU ∗ = U ∗U = I), then U = 1. 5.3. Let G denote the set of invertible elements in a unital Banach algebra. Show that the map of G into G defined by A → A−1 is continuous. 5.4. Suppose that F : Ω → A is a function defined on an open set Ω ⊆ C and taking values in a Banach space A . Show that if f is strongly analytic in Ω , then it is weakly analytic (as defined in Section 5.2). 5.5. Recall that for T ∈ B(H ), the operator T − λ I is invertible if and only if T − λ I is bounded below and has dense range. So one way for λ to get into the spectrum of T is for T − λ I to not be bounded below, meaning that there are unit vectors hn with (T − λ I)hn  → 0. A point λ with this property is said to be an approximate eigenvalue of T ; the set of all approximate eigenvalues of T is called the approximate point spectrum of T . Show (a) Every eigenvalue of T is in the approximate point spectrum of T . (b) The approximate point spectrum of T is a closed set (show its complement is open). (c) Show that if Tn is invertible for all n and Tn → T where T is not invertible, then 0 is an approximate eigenvalue of T . Hints: Explain why it suffices to show that if the range of T is not dense, then there are unit vectors hn with T hn  → 0. Then assume that the range of T is not dense and find a nonzero vector h with h ⊥ ran T (why must such an h exist?). Consider hn = Tn−1 h/Tn−1 h.

5.9 Exercises

147

(d) If λ is in the boundary of σ (T ), then show that λ is an approximate eigenvalue for T . (e) Extend the result of (d) to the case that T is a bounded linear operator on a Banach space X, with “approximate eigenvalue” defined in the analogous way. 5.6. Suppose that T ∈ B(H ). Show that λ is not an approximate eigenvalue of T if and only if T − λ I has a left inverse. 5.7. Let σap (A) denote the approximate point spectrum for an operator A ∈ B(H ). n (A − λ I) is bounded below on H if and only if A − λ I is (a) Show that Π j=1 j j bounded below for 1 ≤ j ≤ n. (b) Show that for any polynomial p, σap (p(A)) = p(σap (A)).

Does the analogous result, with “approximate point spectrum” replaced by “point spectrum” hold? The point spectrum of A is {λ : ker (A − λ ) is nontrivial}, i.e., the set of eigenvalues of A. 5.8. Suppose that A is a bounded linear operator on a Hilbert space H . Show that if A − λ I does not have dense range in H , then λ is an eigenvalue of A∗ , and conversely, if µ is an eigenvalue of A∗ , then A − µ I does not have dense range. Thus the compression spectrum of A can be described in terms of the eigenvalues of A∗ . 5.9. (An Inversion Spectral Mapping Theorem.) Suppose that A is an invertible operator in B(H ). The goal of this problem is to show that   1 σ (A−1 ) = : λ ∈ σ (A) . λ (a) Show that if A − λ I is not bounded below, then A−1 − λ1 I is not bounded below, and conversely that if A−1 − µ I is not bounded below, then A− µ1 I is not bounded below. Show that the eigenvalues of A−1 are precisely the reciprocals of the eigenvalues of A. (b) Show that A − λ I fails to have dense range in H if and only if A−1 − λ1 I fails to have dense range in H . Exercise 5.8 may be helpful here. Conclude that −1

σ (A ) =



 1 : λ ∈ σ (A) . λ

This result holds more generally for any invertible element in a unital Banach algebra; see for example, p. 204 in [8]. 5.10. Consider the operator on ∞ defined by T (x1 , x2 , . . .) = (λ1 x1 , λ2 x2 , . . .) where (λ1 , λ2 , . . .) is in ∞ . Find σ p (T ), σ (T ) and show that σ (T )\σ p (T ) is the residual spectrum of T .

5 Banach and C∗ -Algebras

148

5.11. Recall from Exercise 2.10 in Chapter 2 that if W is a weighted shift and λ ∈ C satisfies |λ | = 1, then λ W is a weighted shift which is unitarily equivalent to W . Show that weighted shifts have “circularly symmetric” spectra, that is, if µ ∈ σ (W ) and |λ | = 1, then λ µ ∈ σ (W ). 5.12. Recall the Banach space C1 [0, 1] = { f : f is continuously differentiable on [0, 1]} with norm  f ∞ +  f  ∞ . (a) Show that under pointwise multiplication, C1 [0, 1] is a Banach algebra. Is it a C∗ -algebra if we define f ∗ = f ? (b) Let g(x) = x for x ∈ [0, 1]. What is the norm of g in C1 [0, 1]? What is the spectral radius r(g)? (c) Show that for each closed set E ⊆ [0, 1], JE ≡ { f ∈ C1 [0, 1] : f (x) = 0 for x ∈ E} is a closed, two-sided ideal in C1 [0, 1]. (d) Find a closed ideal in C1 [0, 1] which is not of the form JE as in (c). 5.13. Suppose A is a Banach algebra and J is a proper, closed ideal. Show that (A + J )(B + J ) = AB + J is a well-defined multiplication on A /J under which this quotient space becomes a complex algebra. 5.14. Recall that an operator T ∈ B(X), where X is a Banach space, is an isometry if T x = x for all x ∈ X. (a) Show that the spectrum of an isometry T is contained in the unit circle ∂ D if T is invertible. (b) Show that if T is an isometry but is not invertible, then its spectrum is D. Hint: By Exercise 5.5, the boundary of the spectrum is contained in the set of approximate eigenvalues of T . (c) Give an example of a continuous ϕ : [0, 1] → [0, 1] so that the composition operator Cϕ (see Exercise 2.3 in Chapter 2) is an isometry on C[0, 1] and σ (Cϕ ) = D. 5.15.(a) An operator T ∈ B(H ) is said to be nilpotent if T n = 0 for some positive integer n. Show any nilpotent operator has spectrum equal to {0}. (b) Say T is quasinilpotent if σ (T ) = {0}. By (a), every nilpotent operator is quasinilpotent. Show that the operator T : 2 → 2 given by T (x1 , x2 , . . .) = (0, is quasinilpotent.

xn x1 x2 , , . . . , n , . . .) 2 4 2

5.9 Exercises

149

5.16. Consider the Volterra integral operator V acting on L2 ([0, 1], dx) defined by  x

V f (x) =

f (t)dt. 0

(a) Show that for any positive integer n, V n+1 f (x) =

1 n!

 x 0

(x − t)n f (t)dt.

(b) Show that σ (V ) = {0}. 5.17. Let A be the Banach algebra C(T ) in the supremum norm, where T denotes the unit circle ∂ D. Let B be the subalgebra of C(T ) consisting of those f ∈ C(T ) for which there exist polynomials pn in z with pn converging uniformly to f on T . (a) Show that g(z) = z is not in B (but of course it is in A ). (b) Consider the function f (z) = z which is in both A and B. What is σA ( f )? Show that σB ( f ) = D, the closed unit disk. Observe that although σA ( f ) = σB ( f ), the spectral radius of the element f doesn’t change in passing from A to B. 5.18. Suppose A and B are Banach algebras with common identity and B ⊆ A . Show σA (A) ⊆ σB (A) and ∂ σB (A) ⊆ ∂ σA (A), for any A ∈ B. Hint for the second part: Since the first part implies that the interior of σA (A) is contained in the interior of σB (A) for any A in B, argue first that it suffices to show that if λ ∈ ∂ σB (A), then λ ∈ σA (A). 5.19. Suppose that A is a C∗ -algebra with unit IA , B is a C∗ -algebra with unit IB and ρ : A → B is a ∗-homomorphism with ρ (IA ) = IB . Prove the following: (a) For every A ∈ A , σ (ρ (A)) ⊆ σ (A), and hence r(ρ (A)) ≤ r(A). (b) For every A ∈ A , ||ρ (A) ≤ A. (c) If ρ is a ∗-isomorphism, then ρ is an isometry. 5.20. Find the norm of the operator on B(H ) where H = C2 which is given by the matrix   ab . cd Give your answer in terms of the numbers S = |a|2 + |b|2 + |c|2 + |d|2 and D = ad − bc. 5.21. Suppose that  · 1 and  · 2 are two norms on a ∗-algebra A , each of which make A into a C∗ -algebra. Show  · 1 =  · 2 . 5.22. Show that the noncommutative unital Banach algebra Mn (C) of all n × n matrices with complex entries has no nontrivial two-sided ideals. (Hints: Take any nonzero matrix A. Show that by multiplying A on the left and right by the appropriate sequence of matrices you can isolate any entry of A and move it anywhere you want. Recall that the elementary row and column operations of interchanging two rows or two columns of A can be obtained by multiplying A by the appropriate elementary matrix.)

5 Banach and C∗ -Algebras

150

5.23. Suppose that J is a closed two-sided ideal in B(H ), for H a Hilbert space. The goal of this problem is to show that either J = {0}, or J contains K (H ), the ideal of compact operators on H . (Compare this with the statement in Exercise 5.22.) (a) Suppose T is a nonzero operator in J . Find vectors f0 , f1 with T f0 = f1 and f1 = 0. Show that if g0 , g1 are any pair of nonzero vectors in H , then the rank one operator S defined by f , g0 g1 Sf = g0 2 is in J . By Exercise 2.6 in Chapter 2, this will show that J contains all rank 1 operators. Hint: Let f , g0 f0 Af = g0 2 and Bf =

f , f1 g1  f1 2

and compute BTA. (b) Apply Exercise 4.9 of Chapter 4 to show that J contains all finite rank operators. 5.24. Let (X, Ω , µ ) be a σ -finite measure space and suppose ϕ ∈ L∞ (µ ). Define Mϕ on L2 (µ ) by Mϕ ( f ) = ϕ f , so that Mϕ is the multiplication operator with symbol ϕ . Recall that Mϕ is a bounded linear operator on L2 (µ ) with Mϕ  = ϕ ∞ . (a) Show that Mϕ is normal, with Mϕ∗ = Mϕ . (b) Show that ϕ → Mϕ is a ∗- homomorphism from L∞ (µ ) into B(L2 (µ )). (c) Show that the eigenvalues of Mϕ are the complex numbers λ for which ϕ −1 ({λ }) has positive measure, and that σ (Mϕ ) is the essential range of ϕ . The essential range of ϕ is defined as: {w ∈ C : µ {x : | f (x) − w| < ε } > 0 for all ε > 0}. (d) Show directly that any closed set in C that contains the range of ϕ must contain the essential range of ϕ . (e) Suppose f is a continuous function on σ (Mϕ ). Identify f (Mϕ ) in the continuous functional calculus. (Hint: Make a guess, and use the uniqueness statement for the functional calculus to prove your guess correct.) 5.25. Suppose that A is a self-adjoint operator in B(H ) for some Hilbert space H . Show that if ker (A − tI) = {0} and A − tI has closed range for some real number t, then the range of A − tI is H . Conclude that a self-adjoint operator has no residual spectrum. 5.26. Suppose S is a set and τ1 , τ2 are topologies on S with τ1 weaker than τ2 . For an arbitrary set A in S, how does the closure of A relative to τ1 compare to the closure

5.9 Exercises

151

of A relative to τ2 ? Is it easier for a set to be compact in the τ1 -topology or the τ2 topology? Is it easier for a sequence (or net) to converge in the τ1 -topology or the τ2 -topology? 5.27. Prove Theorem 5.40 by modifying the proof of Proposition 5.32. 5.28. This problem explains why we require Y to be a vector space in defining the Y -weak topology. (a) Suppose that X is a vector space and ϕ1 , ϕ2 , . . . , ϕn are linear maps from X into C. Let ϕ be a linear map from X into C. Show that ϕ is in the linear span of {ϕ1 , ϕ2 , . . . , ϕn } if and only if ker ϕ1 ∩ ker ϕ2 ∩ · · · ∩ ker ϕn ⊆ ker ϕ . (b) Suppose that Y is a vector space of linear functionals on X that separates the points of X. Show that a linear functional ϕ on X is continuous with respect to the Y -weak topology if and only if ϕ is in Y . 5.29. Suppose A is an infinite-dimensional Banach space and A ∗ is its dual space. (a) Show that a neighborhood basis of 0 in the weak* topology is given by the collection of sets OA1 ,..., An ≡ {ϕ ∈ A ∗ : |ϕ (A j )| < 1, 1 ≤ j ≤ n}, where n is a positive integer, and A1 , A2 , . . . , An are in A . To do this, you need to show that OA1 ,..., An is a weak* open set containing 0, and for any weak* open set O containing 0, there is some positive integer n, and points A j ∈ A with 0 ∈ OA1 ,..., An ⊆ O. (b) Let ϕ0 be in A ∗ . In the weak* topology, a sub-basic neighborhood of ϕ0 has the form {ϕ ∈ A ∗ : |ϕ (A) − ϕ0 (A)| < ε }, where ε > 0 and A is fixed in A . Basic neighborhoods of ϕ0 are finite intersections of sub-basic neighborhoods: N = {ϕ ∈ A ∗ : |ϕ (A j ) − ϕ0 (A j )| < ε j , 1 ≤ j ≤ n} where A j ∈ A . Show that these are always unbounded sets. Hint: Look at a subbasic neighborhood and first suppose ϕ0 (A) = 0, so that the sub-basic neighborhood contains {ϕ : ϕ (A) = 0}, a subspace of A ∗ . (c) Show that the open unit ball in the norm of A ∗ is not weak* open. 5.30. Let X = [0, 1] and put a topology on X by declaring the open sets to be the empty set and those subsets of X whose complement is at most countable. Consider the set A = [0, 1). Show that the closure of A is [0, 1], so that in particular 1 lies in the closure of A. Show that there is no sequence {xn } of points in [0, 1) that converges to 1.

5 Banach and C∗ -Algebras

152

5.31.(a) Suppose S is a Hausdorff topological space. Show that a net in S converges to at most one point; i.e., if xα → x and xα → y then x = y. (b) Give a proof of the “if” direction of Theorem 5.39. 5.32. Let H be a Hilbert space with orthonormal basis {en }∞ 0 and consider the set E = {em + men : 0 ≤ m < n, n = 1, 2, 3, . . .}. Note that E is countable. Show that 0 is in the weak closure of E, but there is no sequence xn ∈ E with xn → 0 weakly. 5.33. Prove the following statement used in Theorem 5.46: If X and Y are homeomorphic compact Hausdorff spaces, then C(X) and C(Y ) are ∗-isomorphic unital C∗ -algebras in a natural way. 5.34. Let A be a unital commutative Banach algebra, and suppose A, B ∈ A . Show that r(A + B) ≤ r(A) + r(B) and r(AB) ≤ r(A)r(B). (Hint: Use the Gelfand transform.) Show the same result holds if A is not assumed to be commutative, provided AB = BA. Show the result fails in general (look in M2 (C)). 5.35. Suppose A is a unital C∗ -algebra, and A ∈ A is a normal element. Show that A is unitary if and only if σ (A) ⊆ ∂ D, the unit circle in the complex plane. Show that A2 = A if and only if σ (A) ⊆ {0, 1}. 5.36. Suppose that N is a normal operator in B(H ) for some Hilbert space H . If λ is in C\σ (N), show that (N − λ I)−1  = dist (σ (N), λ )−1 where dist (σ (N), λ ) is the distance from σ (N) to λ . 5.37. Let W denote the Wiener algebra. Show that the Gelfand transform Γ : W → C(MW ) is not isometric, and indeed is not even bounded below. Thus W cannot be made into a C∗ -algebra (for example, by defining f ∗ = f ). 5.38. The “one-sided Wiener algebra” W+ is defined to be the set of all f in the Wiener algebra W of the form f (eiθ ) =



∑ an einθ .

n=0

(a) Show that W+ is a closed subalgebra of W . inθ ∈ W to (b) For λ in the closed unit disk D, show that ϕλ taking ∑∞ + n=0 an e ∞ n ∑n=0 an λ is a multiplicative linear functional on W+ . (c) Show that MW+ = {ϕλ : λ ∈ D} and the map λ → ϕλ is a homeomorphism of D and MW+ , the latter being equipped with the weak* topology. (d) If f ∈ W+ , describe the spectra σ+ ( f ) and σ ( f ) of f as an element of W+ and W , respectively.

5.9 Exercises

153

5.39. Consider the Banach space 1 (N0 ) of sequences {xn }∞ n=0 in the norm {xn } =



∑ |xn |.

n=0

For {xn } and {yn } in 1 (N0 ), define the convolution {xn } ∗ {yn } to be the sequence {zn } defined by n

zn =

∑ xk yn−k

k=0

for n = 0, 1, 2, . . .. (a) Show that 1 (N0 ) becomes a commutative unital Banach algebra under the convolution product. (b) Show that as a Banach algebra, 1 (N0 ) is isometrically isomorphic to the onesided Wiener algebra W+ defined in Exercise 5.38. (c) For 0 < a < 1, set x = {an } = (1, a, a2 , a3 , . . .). Find σ (x), the spectrum of x in 1 (N0 ). 5.40. Consider the Banach space c of convergent sequences, as defined in Exercise 3.20 of Chapter 3. Define a product and involution respectively on c by {xn } · {yn } = {xn yn } and

{xn }∗ = {xn }.

This makes c a commutative unital C∗ -algebra, with unit I = (1, 1, . . .). Show that its maximal ideal space is Mc = {ϕk : k ∈ N} ∪ {ϕ∞ } where ϕk ({xn }) = xk for k ∈ N and ϕ∞ ({xn }) = limn→∞ xn . Note that the map k → ϕk is a homeomorphism of N onto its range in Mc , and clearly ϕk → ϕ∞ weak* as k → ∞. Thus Mc is naturally homeomorphic to the onepoint compactification (see [33], p.183) of N. 5.41. Consider ∞ as a commutative unital C∗ -algebra with product {xn } · {yn } = {xn yn } and involution {xn }∗ = {xn }. The cluster set at infinity of x = {xn } ∈ ∞ is the set Cl∞ (x) ≡ {λ ∈ C : for every ε > 0 and N ∈ N, there exists n ≥ N with |xn − λ | < ε } = {λ ∈ C : there exists a subsequence {xnk } of {xn } with xnk → λ as k → ∞}. Show that for any x = {xn } ∈ ∞ ,

5 Banach and C∗ -Algebras

154

σ (x) = {xn : n ∈ N} ∪Cl∞ (x) = {xn : n ∈ N}. 5.42. Consider ∞ as a unital C∗ -algebra as in the previous exercise. The fiber at infinity is the subset X∞ ⊆ M∞ defined by X∞ = {ϕ ∈ M∞ : ϕ (x) = lim xn for every x = {xn } ∈ c}. n→∞

(a) For n ∈ N, let en = (0, . . . , 0, 1, 0 . . .) with a 1 in the nth position and 0’s elsewhere. Show that X∞ = {ϕ ∈ M∞ : ϕ (en ) = 0 for all n ∈ N} = {ϕ ∈ M∞ : c0 ⊆ ker ϕ }. (b) Suppose that for each k ∈ N, ϕk ∈ M∞ is defined by ϕk (x) = xk for x = {xn } ∈ ∞ . Show that M∞ = {ϕk : k ∈ N} ∪ X∞ . (c) With the terminology of the last exercise, argue that for any x ∈ ∞ , Cl∞ (x) = {x( ˆ ϕ ) : ϕ ∈ X∞ }. (d) Let x ∈ ∞ . Show that x ∈ c if and only if xˆ is constant on X∞ . (e) Show that {ϕk : k ∈ N} is dense in M∞ in the weak* topology. (f) Let ϕ ∈ X∞ . Show that (the assertion of (e) notwithstanding) there is no subsequence {ϕnk } of {ϕn } with ϕnk → ϕ weak*. 5.43. Can a Banach limit on ∞ (see Exercise 3.9) be multiplicative? 5.44. Recall that in a unital C∗ -algebra A , the positive elements of A are defined to be those self-adjoint A ∈ A with σ (A) ⊆ [0, ∞). Denote the collection of positive elements A+ and set A− = {A : −A ∈ A+ }. (a) Show that A+ ∩ A− = {0}. (b) Follow the outline below to show that every self-adjoint A in a C∗ -algebra A can be written in the form A = A+ − A− , where A+ and A− are both positive elements of A and A+ A− = A− A+ = 0. Outline: Note that the identity function h(t) = t on the real line can be written as f − g, where f (t) = max(0,t) and g(t) = − min(0,t). Observe that f , g are continuous nonnegative functions on the real line with f g = 0. For A self-adjoint, set A+ = f (A) and A− = g(A) as given by the functional calculus. Check that A+ and A− have the desired properties. 5.45. Suppose that 0 ≤ A ≤ B for self-adjoint elements A, B in a C∗ algebra. (a) Show that B ≤ BI. Hint: Consider C∗ (B) ∼ = C(σ (B)) where B corresponds to the identity function on the spectrum of B. Use the functional calculus, with the function f (x) = B − x on σ (B).

5.9 Exercises

155

(b) Show A ≤ B. Hint: Consider C∗ (A) ∼ = C(σ (A)) with A corresponding to the identity function on the spectrum of A. Use the functional calculus with the function f (x) = B − x. (c) Show that 0 ≤ A ≤ B need not imply A2 ≤ B2 by considering   10 X= 00 1

and Y=

1 2 2 1 1 2 2

 .

Show 0 ≤ X ≤ X +Y . Is X 2 ≤ (X +Y )2 ? (d) Show that if 0 ≤ A ≤ B and A and B commute, then An ≤ Bn for every positive integer n. More generally, show that if there are positive elements C j , 1 ≤ j ≤ k, with 0 ≤ A ≤ C1 ≤ C2 ≤ · · · ≤ Ck ≤ B so that any two neighbors in this list commute, then An ≤ Bn for any positive integer n. 5.46. Suppose that P and Q are orthogonal projections onto closed subspaces M and N in H , respectively. Show that P ≥ Q if and only if N ⊆ M. 5.47. Let H be a Hilbert space. An operator T in B(H ) is said to be a contraction if T  ≤ 1. (a) Show that T is a contraction if and only if I − T ∗ T ≥ 0. (b) Suppose that A and B are bounded linear operators on H with B invertible. Show that AB−1 is a contraction if and only if A∗ A ≤ B∗ B. 5.48. What’s wrong with the following “proof” that for an arbitrary element B of a unital C∗ -algebra A , the element B∗ B is positive: Let A = B∗ B. Clearly A is selfadjoint. Using the Gelfand transform we have

Γ (A) = Γ (B∗ B) = Γ (B∗ )Γ (B) = Γ (B)Γ (B) = |Γ (B)|2 ≥ 0 so that σ (A) = σ (Γ (A)) = range |Γ (B)|2 ⊆ [0, ∞).

Chapter 6

The Spectral Theorem

Most students of mathematics learn quite early and most mathematicians remember till quite late that every Hermitian matrix (and in particular every real symmetric matrix) may be put into diagonal form.... The spectral theorem is widely and correctly regarded as the generalization of this assertion to operators on Hilbert space. P. Halmos ([15], p. 241).

The literature of operator theory has a variety of dissimilar looking statements that get called “the spectral theorem.” At their simplest, they describe either compact self-adjoint operators (as we did in Section 4.3) or compact normal operators on a Hilbert space. In either of these cases, a description of the operator connected to eigenvectors is still possible. In this chapter, we want to move to the more general case of bounded normal operators on a Hilbert space H —operators that need not have any eigenvectors.

6.1 Normal Operators Are Multiplication Operators We will begin with a formulation of the spectral theorem that says that bounded normal operators are “multiplication operators” when viewed in an appropriate way. This will give a statement of the spectral theorem which is particularly easy to digest and remember. Throughout this chapter we will work with separable Hilbert spaces. To motivate what will become the statement of the spectral theorem it will be convenient to first look at the finite-dimensional situation. Here the usual formulation of the spectral theorem says that if A is a self-adjoint, or more generally a normal, n × n matrix with complex entries, then Cn has an orthonormal basis consisting of eigenvectors of A. If we denote such an orthonormal basis by {v j }n1 and suppose that Av j = α j v j , then (α1 , α2 , . . . , αn ) is an n-tuple of complex numbers. Think of this n-tuple as a function ϕ in L∞ (X, µ ), where X ≡ {1, 2, . . . , n}, µ is counting measure on the subsets of X, and ϕ (k) = αk . Let W : L2 (X, µ ) → H = Cn be defined by W f = f (1)v1 + f (2)v2 + · · · + f (n)vn and observe that W is a unitary map of L2 (X, µ ) onto Cn . Now suppose that Mϕ is the operator of multiplication by ϕ acting on L2 (X, µ ), and that f is in L2 (X, µ ). B.D. MacCluer, Elementary Functional Analysis, DOI 10.1007/978-0-387-85529-5 6, c Springer Science+Business Media, LLC 2009 

157

158

6 The Spectral Theorem

Since ϕ (k) is by definition the eigenvalue αk for A, we have W Mϕ f = W (ϕ f ) = ϕ (1) f (1)v1 + · · · ϕ (n) f (n)vn = f (1)α1 v1 + · · · + f (n)αn vn = A( f (1)v1 + · · · + f (n)vn ) = AW f . It follows that Mϕ = W −1 AW. This argument can be generalized to any normal T ∈ B(H ), provided T has enough eigenvectors to form an orthonormal basis for H ; see Exercise 6.1. The problem, of course, is that T need not have any eigenvectors. Nevertheless, this point of view suggests we focus on the following concept. Definition 6.1. A bounded linear operator A on a separable Hilbert space H is unitarily equivalent to a multiplication if there is a σ -finite measure space (X, µ ), a function ϕ ∈ L∞ (X, µ ), and a unitary W : L2 (X, µ ) → H such that W Mϕ = AW, where Mϕ denotes the operator of multiplication by ϕ on the (necessarily separable) Hilbert space L2 (X, µ ). In the finite-dimensional example just described, X is {1, 2, . . . , n} and µ is counting measure, so that the norms on L2 (X, µ ) and L∞ (X, µ ) are given by   f 2 =

n

1/2

∑ | f ( j)|

2

j=1

and ϕ ∞ = max |ϕ ( j)|, 1≤ j≤n

respectively. The symbol of the multiplication operator here is the L∞ (X) function ϕ. Notice that only normal operators can be unitarily equivalent to a multiplication, since all multiplication operators are normal. The spectral theorem says the converse: Theorem 6.2 (Spectral Theorem, Multiplication Version). Every normal operator on a separable Hilbert space is unitarily equivalent to a multiplication operator. The goal of this section is to prove Theorem 6.2. Our presentation is influenced by that in [2] and in [15], from which the quote that introduces this chapter is taken. We will first prove Theorem 6.2 under an additional hypothesis, involving the notion of cyclicity, a concept that was briefly explored in Exercise 2.22 of Chapter 2. Definition 6.3. A vector h in H is a cyclic vector for the operator A in B(H ) if (finite) linear combinations of the vectors in the orbit of h, namely

6.1 Normal Operators Are Multiplication Operators

159

{h, Ah, A2 h, A3 h, . . .}, are dense in H . Equivalently, h is a cyclic vector for A if {p(A)h : p is a polynomial} is dense in H . As an example, note that the forward shift S on 2 (N) has cyclic vector h = (1, 0, 0, . . .) (among others). Not every operator has a cyclic vector (the identity operator on a space of dimension greater than one is an easy example). The next example, which is generalized in Exercise 6.2, shows how, for a diagonal matrix, an eigenvalue of multiplicity greater than 1 prohibits cyclicity. Example 6.4. In this example we look at the linear operator on C3 given by the matrix ⎡ ⎤ 100 A = ⎣ 0 1 0 ⎦. 002 Note that for any polynomial p and any vector h ∈ C3 , the first two components of p(A)h agree. Is there a choice of h so that h is cyclic for A? Fix a column vector h = (h1 , h2 , h3 )t . First observe that no h j , j = 1, 2, 3 could be zero if h is a cyclic vector, since the corresponding component of p(A)h would be zero for all polynomials p. Next, the necessarily nonzero vector (h2 , −h1 , 0)t is orthogonal to p(A)h for all p. Thus {p(A)h : p a polynomial} is not dense in C3 , and we conclude that A has no cyclic vector. One can rephrase the invariant subspace problem in terms of cyclic vectors: Since the closure of the span of {h, Ah, A2 h, . . .} is clearly an A-invariant subspace, and the smallest closed A-invariant subspace containing a given vector h must contain the closure of the span of its orbit under A, the question becomes whether every bounded linear operator has a nonzero, noncyclic vector. Enflo’s construction of a Banach space operator with no invariant subspace at its heart proceeds by constructing a Banach space on which a simple multiplication operator has no nonzero, noncyclic vector. Recall that when A is a normal operator in B(H ), the C∗ -algebra C∗ (A) is commutative, and that it is the closure of the polynomials in A and A∗ . We will say that C∗ (A) has a cyclic vector h if {Bh : B ∈ C∗ (A)} is dense in H ; that is, if {p(A, A∗ )h : p = p(z, w) is a two-variable polynomial} is dense in H . Note that if A has a cyclic vector, then trivially so does C∗ (A), and for self-adjoint A, A has cyclic vector h if and only if C∗ (A) has cyclic vector h. Most of the work in proving the spectral theorem will be done in verifying a reduced form of the result, which we state next.

160

6 The Spectral Theorem

Theorem 6.5. Let A be a normal operator in B(H ), where H is a separable Hilbert space. If the C∗ -algebra C∗ (A) has a cyclic vector, then the operator A is unitarily equivalent to a multiplication. Before proceeding to the proof of Theorem 6.5 we make a few comments about what is to be done. We will need to construct a σ -finite measure space (X, µ ) and a unitary operator W : L2 (X, µ ) → H such that W −1 AW is a multiplication operator on L2 (X, µ ). Are there any likely candidates for this measure space, given that our “starting data” is the normal operator A? Based on our previous experience (for example, Theorem 5.46 and the discussion preceding it), σ (A) might be a reasonable candidate for X. It’s less clear where the measure µ is to come from; we will see that the Riesz–Markov theorem, as described in Exercise 3.7 in Chapter 3 and Section A.5 of the Appendix, will produce the measure µ . Moreover, we will see that the multiplication operator we construct in the proof of Theorem 6.5 can be taken to be Mz , the operator of multiplication by the identity function, acting on L2 (X, µ ) for X = σ (A). The first line of the proof of Theorem 6.5 will say “fix a cyclic vector h in H .” The fact that h is cyclic for C∗ (A) only plays a role at the very end of the proof, so the initial steps in the proof hold for any fixed vector h in H . The reader may find it helpful to delay including the cyclicity hypothesis until the end, where its role becomes transparent. Proof (Theorem 6.5). Fix a cyclic vector h in H . Let X = σ (A), the spectrum of A, which is a compact subset of C. We know by Theorem 5.46 that there is a unique isometric ∗-isomorphism γ between C∗ (A) and C(σ (A)) sending A to z, the identity function on σ (A). Properties of this ∗-isomorphism are indicated below; here p is a polynomial in two variables and f , g denote arbitrary functions in C(σ (A)). γ C∗ (A) −→ C(σ (A)) ←− γ −1

I ←− −→ 1

A ←− −→ z A∗ ←− −→ z p(A, A∗ ) ←− −→ p(z, z) f (A) ←− −→ f f (A)∗ ←− −→ f f (A)g(A) ←− −→ f g f (A) + g(A) ←− −→ f + g Define Λ on C(X) = C(σ (A)) by

Λ ( f ) = f (A)h, h ,

6.1 Normal Operators Are Multiplication Operators

161

where f (A) is defined, as above, by the continuous functional calculus and h is our chosen fixed vector. We claim that Λ is a bounded, positive linear functional on C(σ (A)). Linearity follows from the fact that f (A) + g(A) = ( f + g)(A) and (α f )(A) = α f (A) for any scalar α . Since γ is an isometry, |Λ ( f )| = | f (A)h, h | ≤  f (A)h h ≤  f (A) h2 =  f ∞ h2 , so Λ is bounded, with norm at most h2 . Positivity of Λ is verified as follows. A nonnegative function f in C(σ (A)) can be written as f = g2 for some nonnegative g in C(σ (A)). For such f ,

Λ ( f ) = Λ (g2 ) = g2 (A)h, h = g(A)g(A)h, h = g(A)h, g(A)h = g(A)h2 ≥ 0, where we have used the fact that g(A) is self-adjoint in B(H ), since g is self-adjoint (real-valued) in C(σ (A)). Now we invoke the Riesz–Markov theorem for C(X) (see Section A.5 of the Appendix) and conclude that the positive linear functional Λ is given by integration against a unique positive finite Borel measure µ on X = σ (A): f (A)h, h =

 σ (A)

f dµ

for all f ∈ C(σ (A)). The µ -measure of σ (A) is equal to the norm of the linear functional Λ . Next we want a unitary map W : L2 (X, µ ) → H so that W −1 AW is a multiplication operator. We will first define W on C(σ (A)), instead of all of L2 (σ (A), µ ). For f continuous on σ (A), set W f = f (A)h. It is easy to check that W is linear, using the linearity of the ∗-isomorphism γ . Moreover, for f and g continuous on σ (A), W f ,W g H = f (A)h, g(A)h H = g(A)∗ f (A)h, h H = (g f )(A)h, h H = Λ (g f ) 

=

σ (A)

g f dµ

= f , g L2 (X,µ ) . This calculation shows that W is a linear isometry of C(X), equipped with the L2 (X, µ ) norm, into H . Since C(X) is dense in L2 (X, µ ), we can extend W from C(X) to the Hilbert space L2 (X, µ ) by continuity in a unique way. It is easy to check that this extension, which we continue to denote by W , is still a linear isometry into H . What is the range of W ? Since W is an isometry, it has closed range, and thus the range must be a closed subspace of H which contains { f (A)h : f ∈ C(X)} = {Bh : B ∈ C∗ (A)}.

162

6 The Spectral Theorem

Here the cyclicity hypothesis enters. By assumption, {Bh : B ∈ C∗ (A)} is dense in H , and hence W is a unitary map of L2 (X, µ ) onto H . Finally we verify that W −1 AW is a multiplication operator on L2 (X, µ ). We will actually show that (6.1) W −1 f (A)W = M f for all f in C(X). In particular, this shows that W −1 AW = Mz , where z denotes the identity function on X = σ (A). To verify Equation (6.1), compute, for g ∈ C(X), W M f g = W ( f g) = ( f g)(A)h = f (A)g(A)h = f (A)W g, so that W M f = f (A)W on the dense subset C(X) in L2 (X, µ ), and thus also on all of L2 (X, µ ). This completes the proof. We next want to see how to remove the cyclicity hypothesis from Theorem 6.5. Very roughly, the idea is that if C∗ (A) doesn’t have a cyclic vector, we may write H as a direct sum of subspaces Hn , each of which is invariant under A = C∗ (A) (meaning BHn ⊆ Hn for all B ∈ C∗ (A)) and such that the restriction of A to each piece does have the desired cyclicity property. We then apply the reduced form of the spectral theorem to the pieces and use the resulting measure spaces and unitary maps to build a unitary equivalence of A to a multiplication operator. To carry out the details, we begin by reviewing the notion of the direct sum of Hilbert spaces, and the direct sum of operators (see Exercises 1.31 in Chapter 1 and 2.13 in Chapter 2). Given a finite or countable collection of pairwise orthogonal subspaces Hn of a Hilbert space, we denote the set of all convergent sums ∑ hn with hn ∈ Hn by ∑ ⊕Hn . This is isomorphic to the set of all sequences (or k-tuples in the finite case) (h1 , h2 , . . .) where hn ∈ Hn and ∑ hn 2 < ∞ under the map (h1 , h2 , . . .) → h1 + h2 + · · · . When convenient we will think of ∑ ⊕Hn as this set of sequences (or k-tuples) with inner product h, g = ∑ hn , gn . Given operators An ∈ B(Hn ) with sup An  < ∞ n

we define A on ∑ ⊕Hn by A(h1 , h2 , . . .) = (A1 h1 , A2 h2 , . . .). It is easy to check that A is bounded on ∑ ⊕Hn , and that A coincides with the supremum above. We call A the direct sum of the operators An , with respect to the decomposition ∑ ⊕Hn , and write

6.1 Normal Operators Are Multiplication Operators

163

A = A1 ⊕ A2 ⊕ · · · = ∑ ⊕An . n

Lemma 6.6. Let A be a normal operator on a separable Hilbert space H . There is a finite or countable collection of nonzero, pairwise orthogonal subspaces Hn of H satisfying (a) (b) (c)

H = H1 ⊕ H 2 ⊕ · · · . Each Hn is invariant under A = C∗ (A), meaning BHn ⊆ Hn for each B ∈ C∗ (A); equivalently, each Hn is a reducing subspace for A. Each Hn contains a vector hn which is cyclic for An ≡ C∗ (A|Hn ); that is, {Bhn : B ∈ C∗ (A|H n )} is dense in Hn .

Proof. We use a Zorn’s lemma argument. Consider the following family F . An element of F is a collection of nonzero pairwise orthogonal, A -invariant closed subspaces Hα of H , each containing a cyclic vector for C∗ (A|Hα ). The collection F is nonempty since if we pick any nonzero vector h in H , the closure of {A h} is an A -invariant closed subspace of H with a cyclic vector. Partially order F by inclusion. Every totally ordered chain τ in F has an upper bound in F , namely the union of all elements in τ . By Zorn’s lemma there is a maximal element {Hα : α ∈ I} in F . Since H is separable and the Hα are pairwise orthogonal, the index set I is finite or countable, and we write our maximal element as {H1 , H2 , . . .}. We are done if we can show that H = H1 ⊕ H2 ⊕ · · · . If not, the direct sum on the right is a proper closed subspace K of H , which is easily seen to be invariant under A and A∗ . Its orthogonal complement, K ⊥ is nonzero and orthogonal to each H1 , H2 , . . .. Pick a nonzero vector in K ⊥ , say ζ , and look at the cyclic subspace it generates, H0 ≡ {A ζ }− ⊆ K ⊥ . Now {H0 , H1 , H2 , . . .} is in F , contradicting the maximality of {H1 , H2 , . . .}.



The next lemma shows that a direct sum of a sequence (or finite collection) of operators, each of which is unitarily equivalent to a multiplication, is itself unitarily equivalent to a multiplication. It uses the notion of the disjoint union of an at most countable collection of sets. This is defined as follows: Given sets Xn let X˜n = Xn × {n}, so that X˜ j ∩ X˜k = 0/ if j = k, and let

164

6 The Spectral Theorem

X = ∪n X˜n . The parameter n in an element (x, n) ∈ X˜n is there merely to make the sets X˜n and X˜m disjoint, even when Xn and Xm have points in common. If, as will be our case, we have measures µn so that (Xn , µn ) is a measure space for each n, we can define a copy µ˜ n of µn on X˜n by µ˜ n (E × {n}) ≡ µn (E) for each µn -measurable subset E of Xn . A measurable set in X is a set F for which F ∩ X˜n is µ˜ n -measurable for each n; such sets form a σ -algebra on X. We define a measure µ on X by

µ (F) = µ˜ 1 (F ∩ X˜1 ) + µ˜ 2 (F ∩ X˜2 ) + · · ·

(6.2)

for each measurable F ⊆ X. If µn is σ -finite for each n, then µ will be σ -finite also. Lemma 6.7. If A1 , A2 , . . . is a finite or countable collection of operators on, respectively, separable Hilbert spaces H1 , H2 , . . ., with each An unitarily equivalent to a multiplication and supn An  < ∞, then ∑ ⊕An , acting on ∑ ⊕Hn , is unitarily equivalent to a multiplication. Proof. We are given the existence of σ -finite measure spaces (Xn , µn ), unitary operators Wn : L2 (Xn , µn ) → Hn , and functions fn in L∞ (Xn , µn ) such that Wn M fn = AnWn for each n. Moreover, An  = M fn  =  fn ∞ , so that supn  fn ∞ is finite. Let X be the disjoint union of the sets Xn , as described above, and define µ on X by Equation (6.2) for all F with the property that F ∩ X˜n is µ˜ n -measurable for each n. It is easy to verify that L2 (X, µ ) is isometrically isomorphic to ∑ ⊕L2 (Xn , µn ) via the map g ∈ L2 (X, µ ) → (g1 , g2 , . . .) where gn ≡ g|X˜n . Set W = ∑ ⊕Wn , where Wn is our unitary map from L2 (Xn , µn ) onto Hn . We leave it to the reader to check that W is a linear surjection from ∑ ⊕L2 (Xn , µn ) ∼ = L2 (X, µ ) onto ∑ ⊕Hn which is isometric, so that W is unitary. Finally, define the C-valued function f on X by f (x, n) = fn (x) for x ∈ Xn . Observe that f is in L∞ (X, µ ) since supn  fn L∞ (Xn ,µn ) < ∞. Thus M f is a bounded linear operator on L2 (X, µ ) and we claim that  W M f = ∑ ⊕A j W, which is the statement that ∑ ⊕A j acting on ∑ ⊕H j is unitarily equivalent to a multiplication. To see this, let g ∈ L2 (X, µ ) with g = (g1 , g2 , . . .) in ∑ ⊕L2 (Xn , µn ) and compute (W M f )g = W ( f g) = W ( f1 g1 , f2 g2 , . . .) = (W1 ( f1 g1 ),W2 ( f2 g2 ), . . .)

6.2 Spectral Measures

165

= (A1W1 g1 , A2W2 g2 , . . .)  = ∑ ⊕A j W g,

as desired.

We are now ready to prove Theorem 6.2, the full version of the spectral theorem, in its “multiplication form.” Proof (Theorem 6.2). Let A ∈ B(H ) be normal and set A = C∗ (A). By Lemma 6.6 we may write H = ∑ ⊕Hn for pairwise orthogonal nonzero Hn , where A Hn ⊆ Hn and each Hn has an An -cyclic vector. By Theorem 6.5, the restriction An of A to Hn is unitarily equivalent to a multiplication. Since A < ∞, supn An  < ∞. Thus by Lemma 6.7, ∑ ⊕An acting on ∑ ⊕Hn is unitarily equivalent to a multiplication, so that A acting on H is also (where we identify H and ∑ ⊕Hn ). Since the measure µ constructed in the proof of Theorem 6.5 is finite, the underlying measure space in Theorem 6.2 is at least σ -finite. In Exercise 6.3 you are asked to show that in fact this measure space can be taken to be finite. We have used the separability assumption on H here; without it we would not know in Lemma 6.6 that the collection of subspaces is at most countable. There is a version of the spectral theorem for nonseparable Hilbert spaces, which follows along similar lines, but requires a bit more care.

6.2 Spectral Measures In this section we explore the “spectral measure” version of the spectral theorem. This is its more classical articulation, but it will require some effort to develop the somewhat peculiar notion of spectral measures. When A is a normal operator on B(H ), for H a separable Hilbert space, we have used the continuous functional calculus to define f (A), for any continuous function f on σ (A), as schematically illustrated by γ B(H ) ⊇ C∗ (A) −→ C(σ (A)) ←− γ −1

A ←− −→ z

f (A) ←− −→ f Next we will show how to use the spectral theorem to extend the continuous functional calculus to define g(A) to be an operator in B(H ), when g is any bounded Borel measurable function on σ (A). By Theorem 6.2, A is unitarily equivalent to Mϕ , the operator of multiplication by ϕ , on L2 (Y, µ ) for some σ -finite measure space Y and some ϕ ∈ L∞ (Y, µ ). Thus there is a unitary operator W : L2 (Y, µ ) → H such that A = W Mϕ W −1 . As a provisional definition, we propose setting

166

6 The Spectral Theorem

g(A) = W Mg◦ϕ W −1 .

(6.3)

(Later we will give another definition, which is independent of the unitary equivalence of A to a multiplication, and see that it agrees with our provisional definition above.) To see that our provisional definition makes sense, we have one issue to address. Can we choose a representative for ϕ whose range is contained in σ (A)? If so, then g ◦ ϕ will be defined. We know that ess range ϕ = σ (Mϕ ) = σ (A) where “ess range ϕ ” denotes the essential range of ϕ (see Exercise 5.24 in Chapter 5). So our question is whether, altering ϕ on a set of µ -measure zero if necessary, we have range ϕ ⊆ σ (A). If λ is not in the essential range of ϕ , there exists ε > 0 so that the inverse image under ϕ of the disk centered at λ with radius ε has µ -measure zero. The complement of σ (A) is an open set in C, and any open cover of C\σ (A) has a countable subcover ([33], p. 191). Thus C\σ (A) may be covered by a countable collection of open disks whose inverse images under ϕ have µ -measure zero, and therefore ϕ −1 (C\σ (A)) has µ -measure zero. Thus we may alter ϕ on a set of measure zero so that its range is contained in σ (A), as desired, and our proposed definition of “g(A)” in Equation (6.3) is justified. This definition of g(A) for g a bounded Borel measurable function on σ (A) is called the “Borel functional calculus,” and we claim that it extends the continuous functional calculus. To verify this, note that when g(z) = z, then Equation (6.3) gives g(A) = A, and when g(z) = z, Equation (6.3) identifies g(A) as W Mϕ W −1 = A∗ . Now apply the uniqueness statement in Theorem 5.46. We leave the details to the reader. A particular instance of the Borel functional calculus comes from choosing g = χS , the characteristic function of a Borel subset S of σ (A). Observe that (χS ◦ ϕ )(x) = 1 if ϕ (x) is in S, and is equal to 0 otherwise. Thus

χS ◦ ϕ = χϕ −1 (S) and we have

χS (A) = W Mχϕ −1 (S) W −1 .

What kind of operator is this? Since any characteristic function χ is an idempotent (meaning χ 2 = χ ), we have [χS (A)]2 = χS (A). Clearly, χS (A) is self-adjoint. Recall (Exercise 2.17 in Chapter 2) that any operator T in B(H ) that is self-adjoint and satisfies T 2 = T is an orthogonal projection (onto its range). Thus the mapping that sends each Borel subset S of σ (A) to χS (A) associates to each such S an orthogonal projection in B(H ). The identity I ∈ B(H ) is associated to the set σ (A), and the zero operator is associated to the empty set. It is easy to see that orthogonal projections onto mutually orthogonal subspaces of H are associated to disjoint sets. To continue this line of investigation further, we

6.2 Spectral Measures

167

make the following definition. Recall that by a measurable space (X, F ) we mean a set X together with a specific σ -algebra F of subsets of X. Definition 6.8. Let (X, F ) be any measurable space. A spectral measure on X is a function E : F → B(H ), where H is a Hilbert space, satisfying the following: (1) (2) (3)

For each S in F , E(S) is an orthogonal projection. E(0) / = 0 and E(X) = I. / then If S1 , S2 are in F , and S1 ∩ S2 = 0, E(S1 )H ⊥ E(S2 )H .

(4)

If {Sk }∞ 1 is a sequence of pairwise disjoint sets from F , then for each h ∈ H n

∑ E(Sk )h → E(∪∞k=1 Sk )h

k=1

as n → ∞. Note that as a special case of (4), if S1 ∩ S2 = 0, / then E(S1 ) + E(S2 ) = E(S1 ∪ S2 ). Typically our interest in spectral measures will be in the case that X is C or a subset of C, and F is the Borel σ -algebra on X. There are some useful “ordinary” measures, defined on the σ -algebra F , which can be created from a spectral measure E; we will see this in Proposition 6.14 below. Example 6.9. To find an easy example of a spectral measure, let (X, F , µ ) be a measure space and set H = L2 (X, µ ). Define E : F → B(H ) by E(S) = MχS , the operator of multiplication by the characteristic function χS , acting on L2 (X, µ ); the function E is a spectral measure on X. The reader is encouraged to check the details verifying properties (1)–(4) in Definition 6.8. Example 6.10. Suppose that T is a diagonal operator on 2 with diagonal sequence {λ1 , λ2 , . . .}. Define E on the Borel subsets of C by setting E(S) to be the diagonal operator with diagonal sequence {α1 , α2 , . . .}, where α j = 1 if λ j is in S and α j = 0 if λ j is not in S. One easily checks that E is a spectral measure. Example 6.11. Suppose that Mϕ is a multiplication operator on L2 (X, µ ), where (X, µ ) is a measure space. The function E : F → B(L2 (X, µ )) given by E(S) = Mχϕ −1 (S) for S a Borel subset of C is a spectral measure on C. Clearly conditions (1) and (2) of / then for any h1 , h2 ∈ L2 (X, µ ) Definition 6.8 hold. For (3) observe that if S1 ∩ S2 = 0, we have (6.4) E(S1 )h1 , E(S2 )h2 = Mχϕ −1 (S ) h1 , Mχϕ −1 (S ) h2 . 1

2

Disjointness of S1 and S2 ensures that χϕ −1 (S1 ) χϕ −1 (S2 ) = 0, so that the inner product in Equation (6.4) is zero. Property (4) in Definition 6.8 is the statement that for each h ∈ L2 (X, µ ),

168

6 The Spectral Theorem

χϕ −1 (∪n Sk ) h → χϕ −1 (∪∞ Sk ) h 1

1

L2 (X, µ )

in as n → ∞. This is easily verified by, say, an appeal to the dominated convergence theorem. Notice that if S is contained in the complement of the range of ϕ , then E(S) = 0 by definition. Sometimes one defines the support of a spectral measure on C to be the complement of the union of all open sets S for which E(S) = 0. In this example, the support of the spectral measure E is the essential range of ϕ , or equivalently the spectrum of the operator Mϕ (Exercise 6.5). We may think of the spectral measure E in this example as being defined on the Borel subsets of σ (Mϕ ), and extended to all Borel subsets of C by setting E(C) = 0 if C ⊆ C\σ (Mϕ ). Example 6.12. Suppose (X, F ) is a measurable space and E : F → B(H ) is a spectral measure. If K is another Hilbert space and W : H → K is unitary, then the formula F(S) = W E(S)W −1 defines a spectral measure F : F → B(K ). The reader is asked to verify the details in Exercise 6.6. Property (4) in Definition 6.8 is sometimes described by saying that ∞

∑ E(Sk ) = E(∪∞k=1 Sk )

k=1

with the convergence of the sum of projections taking place in the strong operator topology. This terminology appeared earlier, in Exercise 2.23 of Chapter 2. Recall from that exercise that a sequence {Tn } of operators on a Hilbert space is said to converge in the strong operator topology to an operator T if for each vector h, Tn h → ∞ T h. The next result serves to describe the operator E(∪∞ k=1 Sk ) = ∑1 E(Sk ). Recall that ∨Mk denotes the closed linear span of the sets Mk , and when the Mk are pairwise orthogonal closed subspaces, ∨Mk = ∑ ⊕Mk (Exercise 1.33 in Chapter 1). Lemma 6.13. Suppose that H is a Hilbert space and that {Ek }∞ k=1 is a sequence of orthogonal projections on H with E j H ⊥ Ek H for all j = k. We have ∞

∑ Ek = E,

k=1

in the sense of strong operator convergence, where E is the projection of H onto ∞  1



Ek H = ∑ ⊕Ek H . 1

Proof. Let Pn = ∑n1 Ek . We want to show that Pn h → Eh for each h ∈ H , where E is defined as in the statement of the lemma. First suppose that h is in the subspace ∞ n 2 ∑∞ 1 ⊕Ek H , so that h = ∑1 hk with hk ∈ Ek H and ∑ hk  < ∞. Since Pn h = ∑1 hk ,

6.2 Spectral Measures

169

we have Pn h → h = Eh as n → ∞. If, on the other hand, h ⊥ Ek H for all k, then Pn h = 0 = Eh. Given any x ∈ H , write x = y+z where y ∈ ∑∞ 1 ⊕Ek H and z ⊥ Ek H for all k. We have Pn x → y = Ex, as desired. By this lemma, we see that in (4) of Definition 6.8, ∑∞ k=1 E(Sk ) is the projection onto ∑∞ 1 ⊕E(Sk )H , and moreover, for each h ∈ H , ∞

2 2 E(∪∞ 1 Sk )h = ∑ E(Sk )h .

(6.5)

1

Since a spectral measure E maps disjoint sets to projections with orthogonal ranges, it follows easily that S1 ⊆ S2 =⇒ E(S1 )H ⊆ E(S2 )H . To see this, write S2 as the disjoint union of S1 and S2 \S1 , so that E(S2 )H = E(S1 )H ⊕ E(S2 \S1 )H ⊇ E(S1 )H . Using this observation we see that for any two measurable sets B1 and B2 , E(B1 ∩ B2 ) = E(B1 )E(B2 ), and so the operators E(B1 ) and E(B2 ) commute. The details of this statement are left to the reader in Exercise 6.7. Since we have E(Sk )H ⊆ E(∪∞ 1 S j )H for each k, and thus ∞

∑ ⊕E(Sk )H

⊆ E(∪∞ 1 Sk )H ,

1

the significance of condition (4) in Definition 6.8 is to demand the reverse containment. The property in condition (4) of Definition 6.8 is certainly reminiscent of the defining property for a (scalar-valued) measure (i.e., countable additivity), and our next result describes constructing ordinary (complex) measures on (X, F ) from spectral measures. Here we return to our general setting of an arbitrary measurable space (X, F ). Proposition 6.14. Let E : F → B(H ) be a spectral measure, as in Definition 6.8. For each h and g in H , define

µh,g (S) = E(S)h, g

(6.6)

for S in F . This is a finite complex measure on (X, F ) and it has total variation µh,g  at most h g. The measure is positive if h = g. Proof. We first show that µh,g is countably additive. Suppose that {Sk } is a sequence of pairwise disjoint sets in F , and let S denote their union. We have

170

6 The Spectral Theorem ∞

µh,g (S) = E(S)h, g = ∑ E(Sk )h, g

1





1

1

= ∑ E(Sk )h, g = ∑ µh,g (Sk ). When h = g the positivity statement is a consequence of the fact that E(S) is a self-adjoint idempotent:

µh,h (S) = E(S)h, h = E(S)h, E(S)h ≥ 0. To verify the statement about the total variation recall what must be done: Given any partition of X into disjoint measurable subsets {Sk }∞ k=1 , we must show that ∞

∑ |µh,g (Sk )| ≤ h g. 1

Now for such a partition, ∞



1

1 ∞

∑ |µh,g (Sk )| = ∑ | E(Sk )h, g | = ∑ | E(Sk )h, E(Sk )g | 1 ∞

≤ ∑ E(Sk )h E(Sk )g 1

 ≤



∑ E(Sk )h2 1

1/2 



1/2

∑ E(Sk )g2 1

= E(X)h E(X)g = h g, where we have used the Cauchy–Schwarz inequality in both H and 2 , and Equation (6.5). In the case of spectral measures E : F → B(H ), where F denotes the Borel subsets of a compact set X in C, the measures µh,g of Proposition 6.14 are necessarily regular measures (meaning the positive measures |µh.g | are regular). This follows, for example, by Theorem 2.18 in [40]. Given a spectral measure E : F → B(H ), we want to define an operator which deserves to be called  f dE, where f is a bounded measurable complex function on (X, F ). We discuss this heuristically first. It seems reasonable to assign to the choice f = χS the projection E(S), so that we would have

6.2 Spectral Measures

171



χS dE = E(S).

Certainly we would want our “integration” to be linear, so when f is a measurable simple function, that is n

f=

∑ ak χSk ,

k=1

our definition should yield 

n

f dE =

∑ ak E(Sk ).

k=1

Notice that with such a choice for f , we would have   n f dE h, g = ∑ ak E(Sk )h, g = k=1 n

=

k=1

∑ ak µh,g (Sk )

k=1



= 

=

n

∑ ak E(Sk )h, g



n



∑ ak χSk

d µh,g

k=1

f d µh,g

for all h and g in H .  This suggests how we should proceed to formally define f dE, for any bounded measurable f , which we do next. Fix such a function f and consider b : H × H → C defined by  b(h, g) ≡

f d µh,g .

Since µh,g is a complex measure, we remind the reader how the integral on the right hand side of this equation is defined. As a consequence of the Radon–Nikodym theorem, for a given complex measure ν on a set X there exists a measurable function m(x) with modulus 1 so that d ν = md|ν |, where |ν | is the total variation measure of ν , and integration with respect to ν is defined to be integration with respect to md|ν |. Thus   f d µh,g =

f m d|µh,g |,

and m is the Radon–Nikodym derivative of µh,g with respect to |µh,g |. Since for each measurable set S, µh,g (S) = E(S)h, g , we see that b is a sesquilinear form on H × H . Moreover, by Proposition 6.14, |b(h, g)| ≤



| f | d|µh,g | ≤  f X µh,g  ≤  f X hH gH ,

172

6 The Spectral Theorem

where  f X denotes the supremum norm of f on X, and so b is bounded. By Theorem 2.11 there exists a unique bounded operator π ( f ) on H with b(h, g) = π ( f )h, g

(6.7)

for all g and h in H , and moreover π ( f ) ≤  f X . 

Since we have π ( f )h, g = for all h and g, we define



(6.8)

f d µh,g

f dE to be the operator π ( f ). In short 

A=

f dE

is the unique operator with Ah, g = for all h, g in H . It follows that





f d µh,g

χS dE = π (χS ) = E(S)

for any set S ∈ F , as desired in our heuristic discussion above. More generally, using the linearity of π as verified in Theorem 6.15 below, we see that for a simple function n

u=

∑ ak χSk

k=1

we have



u dE = π (u) =

n

∑ ak E(Sk ).

k=1

Thus π (u) is a linear combination of projections when u is simple. If the sets Sk are pairwise disjoint, π (u) has a particularly nice geometric form. Disjointness guarantees that the subspaces E(Sk )H and E(Sm )H are orthogonal for k = m. Each subspace E(Sk )H is invariant for the operator π (u), and on this subspace π (u) acts as the scalar ak times the identity. Writing H as the orthogonal direct sum n

H = K ⊕ ∑ ⊕E(Sk )H ,

(6.9)

k=1

where K is defined by this equation, we observe that π (u) acts on K as the zero operator. Thus with regard to the decomposition (6.9) of H , π (u) acts as a block diagonal matrix

6.2 Spectral Measures

173



⎡ ⎢ 0K ⎢ ⎢ a1 I1 ⎢ ⎢ .. ⎢ . ⎢ ⎢ ⎣ an In

⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎦

0

0

where Ik is the identity on E(Sk )H (see Exercise 4.18 in Chapter 4 for the notion of a matrix with operator entries). It is worth noting that given any bounded F -measurable function f on X, π ( f ) may be approximated in operator norm by operators of the form π (u) with u a simple function as above. In view of the inequality (6.8) (and the linearity of π established below in the proof of Theorem 6.15), it’s enough to produce from f and an arbitrary ε > 0 a simple function u with  f − uX < ε . The standard strategy for doing this (variants of which are implemented in the proofs of Theorems 6.15 and 6.21 below) is to choose pairwise disjoint Borel measurable sets R1 , R2 , . . . , Rq in C, having diameters less than ε , each intersecting the range of f , whose union contains the range of f . The sets Ak ≡ f −1 (Rk ), k = 1, 2, . . . , q, will form a partition of X into nonempty pairwise disjoint measurable sets, and the simple function q

u≡

∑ f (xk )χAk ,

k=1

where xk is an arbitrary point in Ak , does the job: π ( f ) − π (u) = π ( f − u) ≤  f − uX < ε . In the next result, we denote the C∗ -algebra of bounded F -measurable functions on X, in the supremum norm, with pointwise operations and involution f ∗ = f , by B(X, F ). Theorem 6.15. The map π : B(X, F ) → B(H ) is a ∗-homomorphism. Proof. We show first that π is linear. Let h and g be in H and suppose f1 and f2 are bounded measurable functions on X and α is a complex scalar. We have π (α f1 + f2 )h, g =





(α f1 + f2 )d µh,g



f1 d µh,g +



f2 d µh,g

= α π ( f1 )h, g + π ( f2 )h, g

= (απ ( f1 ) + π ( f2 ))h, g .

174

6 The Spectral Theorem

This shows that π (α f1 + f2 ) = απ ( f1 ) + π ( f2 ) and thus π is linear. The fact that π (1) = I follows from the computation π (1)h, g = µh,g (X) = Ih, g . Furthermore, for any bounded measurable f , π ( f )h, g =



f d µh,g =



f d µh,g .

But µh,g (S) = g, E(S)h = µg,h (S), so that π ( f )h, g = π ( f )g, h = g, π ( f )∗ h = π ( f )∗ h, g , which says that π ( f ) = π ( f )∗ . It remains to show that π is multiplicative. Let ε > 0 and partition the plane C (using equally spaced horizontal and vertical lines) into a grid of pairwise disjoint, equal-size semiclosed squares of diameter less than ε (to be specific, say that a square includes its west and south edges, but not its east and north edges, and of the corners, only includes the southwest corner). Since our functions f1 and f2 are bounded, only finitely many of these squares will intersect either the range of f1 or the range of f2 ; label them R1 , R2 , . . . Rq , and note that their union contains the ranges of both f1 and f2 . Set Ak = f1−1 (Rk ) and Bk = f2−1 (Rk ) for each k, 1 ≤ k ≤ q, and from the sets Ai ∩ B j , 1 ≤ i, j ≤ q, list only those that are nonempty, denoting them C1 ,C2 , . . . ,Cp . This gives a partition of X into pairwise disjoint pieces. If x, y lie in the same piece Ck , then | f1 (x) − f1 (y)| < ε and | f2 (x) − f2 (y)| < ε . Choose a sampling point xk in each Ck and define the simple functions p

∑ f1 (xk )χCk

u=

k=1

and

p

v=

∑ f2 (xk )χCk .

k=1

Note that  f1 − uX < ε and  f2 − vX < ε . Since the sets Ck are pairwise disjoint, p

uv =

∑ f1 (xk ) f2 (xk )χCk .

k=1

We may easily compute the product π (u)π (v). Since E(C j )H ⊥ E(Ck )H if j = k, we have    p

π (u)π (v) =



j=1

p

f1 (x j )E(C j )

∑ f2 (xk )E(Ck )

k=1

6.2 Spectral Measures

175 p



=

f1 (x j ) f2 (xk )E(C j )E(Ck )

j,k=1 p

=

∑ f1 (xk ) f2 (xk )E(Ck )

k=1

= π (uv). Replace ε by a sequence εn → 0 and u and v by corresponding sequences un and vn . Note that un , vn , and un vn converge uniformly to f1 , f2 , and f1 f2 , respectively. Thus by (6.8)

π ( f1 f2 ) = lim π (un vn ) = lim π (un )π (vn ) = π ( f1 )π ( f2 ) n→∞

n→∞



as desired. Corollary 6.16. For each f ∈ B(X, F ), π ( f ) is a normal operator in B(H ). Proof. We have

π ( f )π ( f )∗ = π ( f )π ( f ) = π ( f f ) = π ( f )∗ π ( f ). The properties of π given in Theorem 6.15 can be written in spectral integral notation as    (α f1 + f2 ) dE = α f1 dE + f2 dE, 



f dE = and

∗ f dE





f1 f2 dE =

,

  f1 dE

 f2 dE

for bounded F -measurable functions f , f1 , and f2 and scalar α . By a representation of a unital C∗ -algebra, we mean a ∗-homomorphism of the algebra into B(H ) for some Hilbert space H which maps I to I. Theorem 6.15 shows that π is a representation of B(X, F ). A representation is termed faithful if it is injective, and by Exercise 5.19 of Chapter 5, any faithful representation is an isometry. Although we do not prove it here, it turns out that every C∗ -algebra has a representation which is an isometry. This says that every C∗ -algebra “is” a subalgebra of B(H ) for some Hilbert space H . We next explore some consequences of the properties of the representation π in Theorem 6.15. If (X, F ) is a measure space and E : F → B(H ) is a spectral measure we say that a set C ∈ F is a carrier for E if E(X\C) = 0. Since X = C ∪ (X\C) and I = E(C) + E(X\C), C is a carrier exactly when π (χC ) = E(C) = I. In Exercise 6.9 you are asked to show that if f1 and f2 are bounded measurable functions on X with f1 (x) = f2 (x) for all x in some carrier C for E, then

176

6 The Spectral Theorem





f1 dE =

f2 dE.

As an application of these ideas, suppose X = C and E is a spectral measure on the Borel sets of C. The identity function z is of course not bounded on C, but if E has a bounded carrier C, so that zχC is a bounded measurable function on C, then we may define   z dE ≡

zχC dE.

We claim that the right-hand side is independent of the choice of bounded carrier C for E. If C1 and C2 are both bounded carriers, then so is C1 ∩ C2 . This observation follows from noting that C\(C1 ∩C2 ) = (C\C1 ) ∪ (C\C2 ) = (C\C1 ) ∪ (C1 \C2 ) and that C\C1 and C1 \C2 are disjoint with E(C\C1 ) = 0 and E(C1 \C2 ) ≤ E(C \C2 ) = 0. Since zχC1 = zχC2 on the carrier C1 ∩ C2 , we apply Exercise 6.9 to conclude that   zχC1 dE = zχC2 dE. Recall from our discussion following Example 6.11 that the support of spectral measure E : F → B(H ) defined on the Borel sets F of C is the complement of the union of all open sets U with E(U) = 0. The support K is a carrier of E. Although carriers of E are not unique and need not be closed, the support K is unique (and of course also closed). Proposition 6.17. Let E be a spectral measure defined on the Borel subsets of C. The following are equivalent: (a) (b)

E has a bounded carrier. The support of E is compact.

The proof is left to the reader as Exercise 6.10. The proof of the next result, which follows easily from the properties of π , is also left to the reader, as Exercise 6.11 Proposition 6.18. Let (X, F ) be a measure space and E : F → B(H ) a spectral measure. If f is a bounded measurable function on X and h is in H , then

  2 



2

f dE h

= | f | d µh,h .

We use Proposition 6.18 in the next result. Proposition 6.19. Let (X, F ), E, and f be as in Proposition 6.18. We have   f dE = E(S0 )H , ker

6.2 Spectral Measures

177

where S0 = {x ∈ X : f (x) = 0}. Proof. Write π ( f ) =



f dE. By Proposition 6.18, π ( f )h2 =



| f |2 d µh,h

for any h ∈ H . Thus h ∈ ker π ( f ) ⇐⇒



| f |2 d µh,h = 0 ⇐⇒ µh,h (X\S0 ) = 0.

But E(X\S0 )h2 = E(X\S0 )h, E(X\S0 )h

= E(X\S0 )h, h

= µh,h (X\S0 ), so h ∈ ker π ( f ) if and only if E(X\S0 )h = 0. Since I = E(X) = E(X\S0 ) + E(S0 ) where the two summands on the right-hand side have orthogonal ranges, h2 = E(S0 )h2 + E(X\S0 )h2 and h ∈ ker π ( f ) ⇐⇒ E(S0 )h2 = h2 ⇐⇒ E(S0 )h = h ⇐⇒ h ∈ E(S0 )H . 

In the next result we investigate the eigenvalues of z dE. Proposition 6.20. Let F be the Borel subsets of C and suppose E : F → B(H ) is  a spectral measure with compact support. Set A = z dE and suppose λ0 ∈ C. We have E({λ0 }) = 0 ⇐⇒ λ0 is an eigenvalue of A. In this case, ker (A − λ0 I) = E({λ0 })H . Proof. Let D be a closed disk in C containing both λ0 and the support of E. Clearly D is a carrier of E. If we set f = zχD , 

A=

f dE = π ( f ).

By Proposition 6.19, ker (A − λ0 I) = ker π ( f − λ0 ) = E(S0 )H ,

178

6 The Spectral Theorem

where S0 = {λ ∈ C : f (λ ) = λ0 }. Since E(S0 ) = E(S0 \D) + E(S0 ∩ D) = 0 + E({λ0 }) we see that ker (A − λ0 I) = E({λ0 })H and in particular, λ0 is an eigenvalue of A if and only if E({λ0 }) = 0. The next result provides the key idea for the spectral measure version of the spectral theorem. It uses the spectral measure of Example 6.11, associated to a multiplication operator. Theorem 6.21. Let (X, µ ) be a σ -finite measure space and let ϕ be in L∞ (X, µ ). Consider the operator Mϕ acting on L2 (X, µ ) and the spectral measure E(S) = Mχϕ −1 (S) for S a Borel set of C. Then E has compact support, and 

Mϕ =

z dE.

Proof. By Exercise 6.5, the support of E coincides with the essential range of ϕ , that is, with σ (Mϕ ), which is compact. Choose a representative of ϕ whose range is contained in the essential range of ϕ . Let ε > 0 be arbitrary. We begin by constructing a finite collection of pairwise disjoint Borel sets in C, each having diameter less than ε and intersecting the range of ϕ , and whose union is a closed set containing the range of ϕ . Any such collection of sets will serve our purposes here, and we indicate one possible procedure for carrying out such a construction. Take a closed square in C which contains the range of ϕ . Cover this square with a grid of closed sub-squares each of diameter less than ε . Let C1 ,C2 , . . . ,Cq be those closed sub-squares which intersect the range of ϕ . Set D1 = C1 . Find the smallest value of k ≤ q, call it k1 , such that Ck1 \D1 intersects the range of ϕ and set D2 = Ck1 \D1 . If no such k1 exists, the process terminates with D1 . If k1 , and hence D2 , do exist, notice that D1 ∩ D2 = 0/ and D1 ∪ D2 = C1 ∪Ck1 . Next, find the smallest k2 > k1 so that Ck2 \(D1 ∪ D2 ) intersects the range of ϕ , and define D3 = Ck2 \(D1 ∪ D2 ). If no such k2 exists, the process terminates with D2 . If k2 (and thus D3 ) exist, the sets D1 , D2 , D3 are pairwise disjoint and D1 ∪ D2 ∪ D3 = C1 ∪Ck1 ∪Ck2 . Continue, so that if sets D1 , D2 , . . . , Dm have been constructed, at the next step we seek the least km with q ≥ km > km−1 so that Ckm \(D1 ∪ · · ·∪Dm ) intersects the range

6.2 Spectral Measures

179

of ϕ ; if no such km exists, the process terminates with Dm . Otherwise, the set Dm+1 is defined by Dm+1 = Ckm \(D1 ∪ · · · ∪ Dm ). Eventually the process must terminate, and the resulting collection D1 , D2 , . . . , D p are pairwise disjoint Borel sets, each having diameter less than ε , with p range ϕ ⊆ ∪k=1 Dk ≡ K.

Since K = D1 ∪ D2 · · · ∪ D p = C1 ∪Ck1 ∪ · · · ∪Ck p−1 is a union of finitely many closed squares from our original collection of Ck , K is a closed set. Define Ak ≡ ϕ −1 (Dk ) for k = 1, 2, . . . , p and note that A1 , A2 , . . . , A p forms a partition of X into pairwise disjoint, nonempty, µ -measurable sets. Select a sampling point xk in Ak for each k. The simple function p

∑ ϕ (xk )χDk

u=

(6.10)

k=1

is a good approximation to the identity function on K: u − zK ≡ sup |u(z) − z| ≤ ε .

(6.11)

z∈K

Write π for the representation associated to E, so that π ( f ) =



f dE. We have

p

π (u) =

∑ ϕ (xk )E(Dk )

k=1

and E(Dk ) = Mχϕ −1 (D ) = MχA . k

k

Since the Ak form a partition of X, p

∑ χAk = 1.

k=1

We see that for any f in L2 (X, µ )

    2

p p

 2



Mϕ − π (u) f = ϕ ∑ χA f − ∑ ϕ (xk )E(Dk ) f

k



k=1 k=1

2

p



= ∑ (ϕ − ϕ (xk ))χAk f .

k=1

180

6 The Spectral Theorem

Since the sets A1 , A2 , . . . , A p partition X and |ϕ − ϕ (xk )| < ε on Ak , the above coincides with  p



k=1 Ak

|ϕ − ϕ (xk )|2 | f |2 d µ ,

which is bounded above by p

ε2 ∑



k=1 Ak

This says that

| f |2 d µ = ε 2  f 2 .



Mϕ − π (u) ≤ ε .

(6.12)

On the other hand, the inequality (6.8) applied to f = u − z, yields π (u) − π (z) = π (u − z) ≤ u − zK ≤ ε since K, a closed set containing the range of ϕ , contains the support of E. Thus Mϕ − π (z) ≤ 2ε , and since ε was arbitrary we conclude that Mϕ = π (z) =



z dE.

We look next at unitarily equivalent spectral measures. Proposition 6.22. Let H and K be Hilbert spaces and suppose that (X, F ) is a measure space. Let E : F → B(H ) and F : F → B(K ) be spectral measures. If there is a unitary operator W : H → K such that F(S) = W E(S)W −1 for all S ∈ F , then    f dE W −1

f dF = W

for all bounded F -measurable functions f on X. Proof. For any bounded measurable function f on X we let

π( f ) =



f dE and ρ ( f ) =



f dF.

Let ε > 0 and choose, as we have done above, a simple function u on X with  f − uX < ε . If

q

u=

∑ ck χAk ,

k=1

6.2 Spectral Measures

181

then

q

q

k=1

k=1

π (u) = so that W π (u)W −1 =

∑ ck E(Ak ) and ρ (u) = ∑ ck F(Ak ) q

∑ ckW E(Ak )W −1 =

k=1

q

∑ ck F(Ak ) = ρ (u).

k=1

Thus W π ( f )W −1 − ρ ( f ) = W (π ( f ) − π (u))W −1 + ρ (u) − ρ ( f ) ≤ W (π ( f ) − π (u))W −1  + ρ (u) − ρ ( f ) = π ( f − u) + ρ ( f − u) ≤ 2 f − uX < 2ε . Since ε is arbitary, ρ ( f ) = W π ( f )W −1 .



The next result is the “spectral measure version” of the spectral theorem. The reader is encouraged to see a concrete application of its statement by working out Exercises 6.13 and 6.14 before proceeding to the proof of the theorem. The intuition behind this theorem is that just as a bounded, Borel measurable function can be approximated by linear combinations of characteristic functions associated to pairwise disjoint sets (i.e., simple functions), bounded normal operators can be approximated by linear combinations of projections with pairwise orthogonal ranges. Theorem 6.23 (Spectral Theorem, Spectral Measure Version). Let A be a normal operator in B(H ), where H is a separable Hilbert space. Let F denote the Borel subsets of σ (A). There is a unique spectral measure F : F → B(H ) with 

A=

z dF.

Moreover, given any continuous function f on σ (A), 

f (A) =

f dF,

where on the left-hand side f (A) is defined by the continuous functional calculus (Theorem 5.46). Proof. We address the existence part of the statement first. By Theorem 6.2 we know that there is a σ -finite measure space (X, ν ), a unitary operator W : L2 (X, ν ) → H , and a ϕ ∈ L∞ (X, ν ) so that A = W Mϕ W −1 . Let E be the spectral measure associated to Mϕ as in Example 6.11. We know from Theorem 6.21 that 

z dE = Mϕ . By Example 6.12, F(S) ≡ W E(S)W −1 is a also spectral measure defined on the Borel subsets of σ (A). We claim that A = zdF. This is immediate from Proposition 6.22, with f taken to be the restriction of z to σ (A). Indeed,

182

6 The Spectral Theorem





z dF = W



z dE W −1 = W Mϕ W −1 = A.

This proves the existence part of the first statement of the theorem; the proof of the uniqueness statement is outlined in Exercise 6.16. For the second statement of the theorem, let f be continuous on σ (A) and set   π ( f ) = f dF. By Theorem 6.15, p(A, A∗ ) = p(z, z)dF for any polynomial p in z and z. Since f is a uniform limit on σ (A) of a sequence of such polynomials, we have  f (A) = f dF

as desired.

For a normal operator A in B(H ), the expression “the spectral measure of A” always refers to the unique measure given in Theorem 6.23. Notice that when E is the spectral measure of A, the statement 

g(A) =

gdE

holds for any function g which is continuous on σ (A). The same identity, with “g(A)” defined by the Borel functional calculus (that is, by Equation (6.3)), holds for any bounded Borel measurable function; see Exercise 6.18. By Proposition 6.20 we know that when H is separable and A is a normal operator in B(H ) with spectral measure E, then the eigenvalues of A are precisely those points λ0 ∈ σ (A) for which E({λ0 }) is nonzero, and moreover E({λ0 })H = ker (A − λ0 I). Let’s interpret this in the context of a compact normal operator T . We know from Theorem 4.31 that in this case the nonzero points of σ (T ) are all eigenvalues, and this set is at most countable, accumulating only at zero if infinite. Denote these nonzero eigenvalues by λ1 , λ2 , . . .. Set Ek = E({λk }), where E is the spectral measure of T . Each Ek is the projection onto ker(T − λk I). Moreover, by Exercise 4.10 in Chapter 4, Ek is finite rank (that is, it is the projection onto a finitedimensional subspace), since a compact operator has finite-dimensional eigenspaces corresponding to its nonzero eigenvalues, and Ek H ⊥ En H for k = n. What is ∑k λk Ek ? Set fk = λk χ{λk } and notice that

∑ fk (z) = z k

on σ (T ) and that fk (T ) = λk Ek . Thus T = ∑ fk (T ) = ∑ λk Ek . k

If the set of nonzero eigenvalues λk is infinite, the sum converges in the norm of B(H ), since compactness of T implies that λk → 0 as k → ∞. Notice that with this application of Theorem 6.23, we have recovered Theorem 4.24, in the case that T

6.3 Exercises

183

is compact and self-adjoint, and provided an extension of Theorem 4.24 to compact normal operators.

6.3 Exercises 6.1. Suppose a normal operator A in B(H ) has enough eigenvectors to provide an orthonormal basis for H : Aek = αk ek where {ek } is an orthonormal basis for H . (a) Check that the sequence (α1 , α2 , α3 , . . .) is in ∞ (N). (b) Define W : 2 → H by W (λ1 , λ2 , . . .) = λ1 e1 + λ2 e2 + · · ·. Check that W is a unitary map of 2 onto H . (c) Define the operator B in B(2 ) by B ≡ W −1 AW . Identify B as a multiplication operator on 2 , where we regard 2 = L2 (N) in the usual way. 6.2. Let H = Cn and let A be the operator on H given by the n × n diagonal matrix with diagonal (a1 , a2 , . . . , an ). Show that A has a cyclic vector if and only if all the diagonal entries are distinct. Hints: If the a j are distinct, consider the vector h in Cn of all 1’s. For the converse, mimic the argument in Example 6.4. 6.3. Show that every normal operator A acting on a separable Hilbert space is unitarily equivalent to a multiplication operator on L2 (X, µ ) for some finite measure space (X, µ ). Hint: If Hn is invariant under C∗ (A) and has cyclic vector hn , then we may assume hn  = 2−n . 6.4. A conjugation on a Hilbert space H is a conjugate linear map C : H → H with C2 = I and Cx,Cy = y, x for all x, y in H . (a) Show that if we fix an orthonormal basis {en } for H and set C(∑ λn en ) = ∑ λn en (provided {λn } ∈ 2 ) then C is a conjugation. (b) If H = L2 (X, µ ) for some σ -finite measure space (X, µ ), show that C( f ) = f is a conjugation. (c) If C is a conjugation, an operator T ∈ B(H ) is called C-symmetric if CT ∗ = TC. Show that the Volterra operator V f (x) = 0x f (t)dt on L2 [0, 1] is C-symmetric for C f (x) = f (1 − x). (d) If N is any normal operator in B(H ), find a conjugation C on H so that CN ∗ = NC. 6.5. Verify the assertion made in Example 6.11 about the support of E. 6.6. Provide the details in Example 6.12.

184

6 The Spectral Theorem

6.7. Let E be a spectral measure on (X, F ). Show that if S1 and S2 are in F , then E(S1 ∩ S2 ) = E(S1 )E(S2 ). 6.8. Suppose that X is a topological space and F is the Borel σ -algebra on X. A spectral measure E on (X, F ) is said to be regular if E(S)H = closed linear span {E(K)H : K ⊆ S is compact} for every S in F . Show that if X = C, then any spectral measure E : F → B(H ) is necessarily regular. Hints: Clearly, closed linear span {E(K)H : K ⊆ S is compact} ⊆ E(S)H . For the reverse inclusion, suppose that h is a nonzero vector in H with h ⊥ E(K)H for all compact subsets K of S. Use the fact that the measure µh,h is automatically regular ([40]) to show that h ⊥ E(S)H . 6.9. Suppose that (X, F ) is a measure space and E : F → B(H ) is a spectral measure. Show that if f1 and f2 are bounded measurable functions on X with f1 (x) = f2 (x) for all x in some carrier C for E, then 



f1 dE =

f2 dE.

6.10. Prove Proposition 6.17. 6.11. Prove Proposition 6.18. 6.12. Let F be the Borel subsets of C and suppose E : F → B(H ) is a spectral  measure with compact support K. Set A = z dE. Show that σ (A) = K. Hint: If λ0 ∈ K, show that A − λ0 I is not bounded below by considering unit vectors in E(Dε (λ0 ))H , where Dε (λ0 ) is the open disk of radius ε centered at λ0 . 6.13. Let T be the diagonal operator with diagonal {λ1 , λ2 , . . .} on 2 . Consider the spectral measure E defined in Example 6.10. Describe concretely the associated measures µh,g , the operator π (z), and verify directly that 

T=

z dE.

6.14. Let A be the normal operator of multiplication by ϕ (x) = x on L2 [0, 1]. Let E : [0, 1] → B(L2 [0, 1]) be the spectral measure of A as in Theorem 6.23. Identify E. 6.15. Suppose E is a spectral measure and suppose that E(S)  commutes with an operator T for each S in F . Show that for any f ∈ B(X, F ), f dE commutes with T. 6.16. Prove the uniqueness statement in Theorem 6.23. One possible outline for the argument is as follows.

6.3 Exercises

185

(a) Your goal is to show that if F : F → B(H ) is the spectral measure constructed in the proof of Theorem 6.23, and F1 : F → B(H ) is a second spectral measure with   z dF = z dF1 1 =µ then F = F1 . Argue that it suffices to show µh,g h,g for all h, g ∈ H , where 1 µh,g (S) = F1 (S)h, g and similarly for µh,g . 1 =µ (b) By polarization, argue that it suffices to show µh,h h,h for all h ∈ H .     1 , as well as the analogous (c) From z dF = z dF1 we obtain zd µh,h = zd µh,h conclusion with z replacing z. We have



p(z, z) d µh,h =



1 p(z, z) d µh,h

for polynomials p. Now invoke the Stone–Weierstrass theorem. 6.17. Suppose A is a normal operator in B(H ) with spectrum X and let E : F → B(H ) be its spectral measure as in Theorem 6.23. (a) If X consists of a single point, show that A is a scalar multiple of the identity. Conclude that every subspace of H is an invariant subspace of A in this case. (b) Show that if X = S1 ∪ S2 where S1 and S2 are disjoint nonempty Borel subsets of X, then E(S1 ) commutes with A. Moreover, the ranges of E(S1 ) and E(S2 ) are invariant subspaces of A with (E(S1 )H )⊥ = E(S2 )H . 6.18. For T any normal operator in B(H ) and g any bounded Borel measurable function on X = σ (T ), we define g(T ) by the Borel functional calculus (Equa tion (6.3)). Show that g(T ) = gdE where E is the unique spectral measure in Theorem 6.23. 6.19. Suppose that A is a normal operator in B(H ) and that λ0 is an isolated point in σ (A). Show that λ0 is an eigenvalue of A. 6.20. Suppose that H is an infinite-dimensional separable Hilbert space and that T ∈ B(H ) is a compact normal operator. Let f be any function in C(σ (T )). Show that f (T ) is compact if and only if f (0) = 0. 6.21. Suppose A is normal on B(H ), so that f (A) is also normal for any f ∈ C(σ (A)). If E1 is the spectral measure of A and E2 is the spectral measure of f (A), show that E2 (S) = E1 ( f −1 (S)) for all Borel sets S in σ ( f (A)). 6.22. Let M and N be closed subspaces of a Hilbert space H . Let P and Q be, respectively, the orthogonal projections of H onto M and N. (a) Show that the sequence {An } of operators given by An = PQPQ · · · PQP

186

6 The Spectral Theorem

(with n Q’s and (n + 1)P’s) converges in the strong operator topology to the orthogonal projection of H onto M ∩ N. Hint: Apply the spectral measure version of the spectral theorem to the operator PQP. (b) The result in part (a) has been called the zig-zag theorem. Draw a picture (as if H were the real Hilbert space R3 and M and N were two-dimensional subspaces) illustrating why.

Appendix A

Real Analysis Topics

It is clear then, that we must partition not [a, b], but rather the interval [α , β ] bounded by the upper and lower bounds of f on [a, b].... Henri Lebesgue, Copenhagen, 1926.

As a way to illustrate the difference between the Riemann integral and the Lebesgue integral, consider the analogy of finding the value of a pile of coins. The “Riemann” way is to take the coins as they appear, adding the value of each piece as you pick it up. By contrast, in the Lebesgue method we start by sorting the coins by type — penny, nickel, dime, . . . — and total the value as 1 · m(A1 ) + 5 · m(A2 ) + 10 · m(A3 ) + · · · where m(A1 ) is the number of pennies, m(A2 ) the number of nickels, and so on.1 Notice how this approach suggests an interest in sums of the form ∑ s j m(A j ) where “m” is some way of measuring the “size” of sets under consideration. Thus before we pursue a definition of the Lebesgue integral we will discuss the notion of measures.

A.1 Measures Definition A.1. A measure space (X, M, µ ) consists of a set X, a collection M of subsets of X, and a function µ : M → [0, ∞] . The collection M is a σ -algebra, that is, it is required to satisfy (1) (2) (3)

X ∈M If A is in M, then so is its complement Ac . If An is in M for n = 1, 2, 3, . . ., then so is ∪∞ n=1 An .

Furthermore, the set function µ must be countably additive: If {An } is a countable collection of pairwise disjoint sets in M, then

1

This analogy was proposed by Henri Lebesgue himself in a 1926 address in Copenhagen in which he discussed the origins of his ideas for his theory of integration. An English translation of this address can be found in [7] or [29].

187

188

A Real Analysis Topics

µ(

∞ 

n=1



An ) =

∑ µ (An ).

n=1

To avoid a trivial situation, we also require the existence of some A ∈ M with µ (A) < ∞. The sets in M are called the (µ -)measurable sets. Conditions (2) and (3) say that the collection of measurable sets is closed under complementation and countable unions. The set function µ , whose domain is the collection of all measurable sets in X, is called a measure on X. When the σ -algebra M or the measure µ is clear from the context, they are often omitted from the notation (X, M, µ ). There are such things as signed measures (taking values in the real line R) and complex measures (taking values in C), but unless we say explicitly otherwise, “measure” will always mean “positive measure” in the sense of Definition A.1. If the values of µ are restricted to [0, ∞) it is called a finite measure. A measure is said to be σ -finite if the underlying set X can be written as a countable union of (measurable) sets each having finite measure. One can easily obtain the following as consequences of Definition A.1 and simple set-theoretic manipulations: (a) (b) (c) (d)

0/ ∈ M, and hence a finite union of measurable sets is measurable. A finite or countable intersection of measurable sets is measurable. µ (0) / = 0. If A and B are measurable sets, with A ⊆ B, then µ (A) ≤ µ (B); this says µ is monotone.

We’ll see shortly why we want the flexibility to have M be a proper subset of the collection P(X) of all subsets of X. Nevertheless, there are important examples of measure spaces where M = P(X), and hence where the requirements (1)–(3) of Definition A.1 are automatically satisfied. Example A.2. Let X = N, the natural numbers, set M = P(N), and let µ assign to each finite subset of N its cardinality, and to each infinite subset of N the value ∞. With the convention that a + ∞ = ∞ + a = ∞ for 0 ≤ a ≤ ∞, verification that µ is countably additive is immediate. This is called counting measure on the positive integers. Notice that the only set with counting measure zero is the empty set. Counting measure on N is not a finite measure, but it is σ -finite. Example A.3. Let X be any set, let M = P(X) and fix an arbitrary point x0 in X. Define  1 if x0 ∈ A µ (A) = (A.1) 0 otherwise for each A ⊆ X. Verification that (X, M, µ ) is a measure space is easy. The measure µ is called the (unit) point mass measure at x0 . Our most important example will be “Lebesgue measure” on the real line R or on an interval [a, b] ⊆ R. The underlying idea is to generalize the notion of “length”

A.1 Measures

189

from intervals to more general sets. That is, we seek a measure space (R, M, m) where m(I) = |I| = d − c whenever I is an interval with endpoints c and d. Furthermore, we want this measure m to be translation invariant, so that m(A + x) = m(A) for every A ∈ M, where A + x ≡ {a + x : a ∈ A} is the translate of A by x ∈ R. Perhaps unexpectedly—and this explains why our definition of a measure space doesn’t require that M = P(X)—it is impossible to do this if we want every subset of R to be measurable. As soon as we ask that the measure of an interval be its length, and require the measure to be translation invariant, there must exist nonmeasurable sets.2 Fortunately, it is possible to satisfy our desired properties with a σ -algebra L of subsets of R that is sufficiently rich to include all open sets in R. In particular, there is a smallest σ -algebra containing all the open sets, called the Borel σ -algebra. In addition to all open sets, the Borel σ -algebra contains all closed sets, any countable union of closed sets, any countable intersection of open sets, and so on. The Lebesgue measurable sets, which we discuss next, form a σ -algebra L which (properly) contains the Borel sets. While the details of the construction of L and of Lebesgue measure on (R, L ) will not be given here, it is easy to give an outline as to how to proceed. For more information the reader is referred to [39], for example. Motivated by the desire to have the measure of an interval be its length, we look at all ways of covering an arbitrary set A ⊆ R by a countable collection of open intervals {In } and define m∗ (A) ∈ [0, ∞] by $ # m∗ (A) = inf



∞ 

n=1

n=1

∑ |In | : A ⊆

In .

This is called the Lebesgue outer measure of A; it is defined for all subsets of R and is translation invariant. It should also be clear the outer measure is monotone, so that A ⊆ B implies that m∗ (A) ≤ m∗ (B). The outer measure of an interval is its length. Outer measure fails to be countably additive, but there is a proper subset L of P(R) which is a σ -algebra containing all open sets, so that the restriction of m∗ to L is countably additive. The subsets of R that belong to L are defined to be those sets A satisfying m∗ (T ) = m∗ (T ∩ A) + m∗ (T ∩ Ac ) for all T ⊆ R, where Ac is the complement of A in R. This definition, which is not Lebesgue’s original one, but is rather due to Carath´eodory, is perhaps not completely transparent. We use T for “test” set; we are testing the additivity of the outer measure of arbitrary sets against A. Some observations follow easily from Carath´eodory’s definition. For example, every set with outer measure zero is Lebesgue measurable (i.e., it belongs to L ), and a set belongs to L if and only if its complement does. With some effort, one shows that the measurable sets form a σ -algebra which contains all intervals, and thus all open sets. 2

For an example of a nonmeasurable set, and an interesting discussion of the role of the axiom of choice in its construction, see [4].

190

A Real Analysis Topics

When A ∈ L , we define its Lebesgue measure m(A) by m(A) = m∗ (A); that is, Lebesgue measure is simply the restriction of outer measure to the Lebesgue measurable sets L . Once we have defined Lebesgue measure on R, we can also consider Lebesgue measure on any interval [a, b], by intersecting measurable sets in R with [a, b]. We’ll reserve the notation m for Lebesgue measure. Sets of measure zero are easy to understand: A subset of R has measure zero precisely when for each positive ε it can be covered by an at most countable collection of intervals the sum of whose lengths is less than ε . Countable sets have measure zero, but there are uncountable sets of measure zero as well.3

A.2 Integration Let’s recall how the Riemann integral of a bounded, real-valued function on the interval [a, b] is defined. Partition the interval into subintervals by means of subdivision points a = x0 < x1 < x2 < · · · < xn = b. For each such partition, we have the upper and lower sums Uf =

n

n

j=1

j=1

∑ M j (x j − x j−1 ) and L f = ∑ m j (x j − x j−1 )

where Mj =

sup

x j−1 ≤x≤x j

f (x)

and

mj =

inf

x j−1 ≤x≤x j

f (x).

We say that f is Riemann integrable on [a, b] if inf(U f ) = sup(L f ), where the infimum and supremum are taken over all partitions of [a, b] as just described. The common value of this infimum and supremum is the Riemann integral of f over [a, b]. In the 1820s Cauchy, who is credited with the first attempt at a rigorous definition of continuity, had considered sums of a similar sort (choosing f (x j−1 ) instead of m j or M j ), but he assumed a priori that f was continuous. By contrast, Riemann as part of the work for his Habilitation degree in 1854, did not suppose f to be continuous, and thus called attention to the question: What functions are (Riemann) integrable? To illustrate some of the nuances of this question, Riemann gave an example of a function whose discontinuities are dense in the the real line, but which is nevertheless Riemann integrable on any finite interval [a, b]. It is easy to see, however, that “too many” discontinuities can cause trouble. The example, 3

The Cantor ternary set provides an example, since its complement in [0, 1] is a collection of disjoint intervals whose lengths sum to 1.

A.2 Integration

191

due to Dirichlet, of the function defined on [0, 1] by  1 if x is irrational f (x) = 0 if x is rational

(A.2)

has U f = 1 and L f = 0 for each partition of [0, 1], and hence f is not Riemann integrable. This function is discontinuous at every point. One can show that a bounded function on [a, b] is Riemann integrable if and only if the set of points at which it fails to be continuous has Lebesgue measure zero.4 By the 1870s Riemann’s theory of integration had become widely known, and had had successful application in a number of areas. Some limitations of Riemann’s method had also come to light, but these were not yet regarded as serious deficiencies. From a modern perspective we can see several problems with Riemann’s theory of integration. Most simply stated, not enough functions are Riemann integrable. There is an incompatibility of Riemann integration and limit processes. Every potential difficulty with the statement   % & lim fn , (A.3) lim fn = n→∞

n→∞

where the integrals are Riemann integrals, can occur. For example, one side of Equation (A.3) may fail to exist, even if the other side is perfectly well-behaved, or both may exist, but they fail to agree. While the Lebesgue integral doesn’t remove all problems with this interchange of limit and integral, we can give several useful conditions under which Equation (A.3) holds; see Theorems A.4 and A.6 below. From a functional analysis perspective, there is another serious deficiency of the Riemann integral. The set of all Riemann integrable functions f : [0, 1] → R in the metric  1

d( f , g) = 0

| f (x) − g(x)|dx

fails to be complete.5 In other words, in this important metric, there are Cauchy sequences of Riemann integrable functions that fail to converge to a Riemann integrable limit. The Lebesgue integral, introduced by Lebesgue in his doctoral thesis of 1902, led to a resolution of this fundamental problem. We now turn to the definition of the Lebesgue integral. Because it does not rely on a partitioning of the domain, the definition can be just as easily made for an arbitrary measure space (X, M, µ ) as for the particular example (R, L , m), and we will do so.

4

A confusion between the measure-theoretic notion of smallness (Lebesgue measure zero) and the topological notion of smallness (nowhere dense, in the language of Section 3.2) muddied some of the initial study of Riemann’s notion of integral. There are variants of the Cantor set which are nowhere dense but have positive Lebesgue measure. 5 Strictly speaking, d is not a metric, since d( f , g) = 0 does not imply that f (x) = g(x) at every x ∈ X. This problem can be easily rectified, though; see the discussion in Section A.3 below.

192

A Real Analysis Topics

Consider first a function s defined on X and taking only finitely many distinct, nonnegative values α1 , α2 , . . . , αn . If for each j = 1, 2 . . . , n, the set A j = s−1 (α j ) is in M, we call s a nonnegative measurable simple function (a simple function in general is one taking only finitely many distinct values). The sum n

∑ α j µ (A j )

j=1

is a value in [0, ∞] (we define α j µ (A j ) = 0 if α j = 0 and µ (A j ) = ∞). We call this the Lebesgue integral of the nonnegative simple function s with respect to µ , and denote it  s dµ. X

Furthermore, for any measurable subset E of X, define  E

s dµ =

n

∑ α j µ (A j ∩ E).

j=1

Notice how we need each A j to be µ -measurable for these definitions to make sense. An arbitrary real-valued function f on X is said to be measurable if f −1 [α , β ) ≡ {x ∈ X : α ≤ f (x) < β } is in M for each α , β ∈ R. There is a fair amount of flexibility in this definition. For example, the half-open intervals [α , β ) can be replaced by open intervals, or closed intervals, or arbitrary open sets, or arbitrary closed sets in R. In measure spaces where all sets are measurable (like that of Example A.2), all functions are measurable. The definition of measurability shows that if we take a bounded, nonnegative, measurable function f on X, we can approximate f by measurable simple functions in the following natural way. Suppose that 0 ≤ f ≤ M on X and let n be a positive integer. As in the quote of Lebesgue which introduces this chapter, we partition [0, M] into nonoverlapping subintervals I j by     1 2 1 I1 = 0, , I2 = , n n n and in general

 Ij =

j−1 j , n n



−1 (I ) for j = 1, 2, . . . , Mn + 1, so that for j ≤ Mn + 1. Define sn on X to be j−1 j n on f sn is a measurable simple function and

A.2 Integration

193

0 ≤ f (x) − sn (x) ≤

1 n

on X.

y = f (x)

Ij

PP i 1  PPI @ @  f −1 (I j ) FIGURE A.1: Constructing the measurable simple function sn With this approximation scheme to aid our intuition, we define the Lebesgue integral of any measurable f : X → [0, ∞] as    f d µ = sup s d µ : s is a simple measurable function and 0 ≤ s ≤ f , X

X

and say that f is Lebesgue integrable if this supremum is finite. We integrate over a measurable subset E of X by defining    f d µ = sup s d µ : s is a simple measurable function and 0 ≤ s ≤ f . E

E

Two monotonicity properties follow readily from these definitions: 0 ≤ f ≤ g =⇒

 A

and 0 ≤ f and A ⊆ B =⇒

f dµ ≤  A

 A

f dµ ≤

g dµ  B

f dµ.

In a general measure space (X, M, µ ) we can always assume that every subset of any set of µ -measure zero belongs to M (see Theorem 1.36 in [40]; we have already observed this property for m). This implies that if a measurable function on X is changed on a set of µ -measure zero, the result is still measurable. Moreover, integrals are not affected by such a change. Thus many results in the theory of Lebesgue integration are stated with the provision “almost everywhere,” meaning, except possibly on a set of measure zero. For an example, see Theorem A.6 below.

194

A Real Analysis Topics

As a simple exercise in using this terminology, the reader is invited to show that for a nonnegative measurable function f and A ∈ M,  A

f d µ = 0 if and only if f = 0 almost everywhere on A.

Lebesgue conceived of his theory of integration as an extension of Riemann’s, and a Riemann integrable function on an interval [a, b] in R will also be Lebesgue integrable, with equality of the integrals. The function in (A.2) provides an example of a measurable simple function on [0, 1] with Lebesgue integral equal to 1 that is not Riemann integrable. We state two theorems about interchange of limit and integral. Both concern sequences of measurable functions on an arbitrary positive measure space (X, M, µ ) and their Lebesgue integrals, and give useful conditions under which a statement in the form of Equation (A.3) holds. Theorem A.4 (Monotone Convergence Theorem). Suppose that { fn } is a sequence of measurable functions with 0 ≤ f1 (x) ≤ f2 (x) ≤ · · · ≤ ∞ for all x ∈ X. If f (x) = limn→∞ fn (x) for each x ∈ X, then f is measurable, and 

lim

n→∞ X

fn d µ =

 X

f dµ.

Notice that monotonicity implies that the limit function f always exists, so long as we permit it to take values in [0, ∞]. As an application of Theorem A.4, we encourage the reader to verify the details of the following example. Example A.5. When µ is counting measure on the set N of natural numbers,  N

f dµ =



∑ f (n)

n=1

for any nonnegative-valued function f on N. Any such function can be realized in a natural way as a monotone increasing limit of simple functions. Our next theorem on the interchange of limit and integral is normally stated for complex-valued functions. Before giving its statement, we need to extend our definition of the Lebesgue integral to this larger class. Any measurable real-valued function f can be written as a difference f + − f− of two nonnegative measurable functions, where f + = max( f , 0) and f− = − min(0, f ). For complex-valued functions f = u + iv, measurability is defined by requiring that the real-valued functions u and v be measurable. When f is measurable, so is its modulus | f |. This means that the definition of Lebesgue integration can be readily extended to certain complex valued functions as follows: Provided X | f |d µ < ∞, define for any µ -measurable set A in M

A.2 Integration

195

 A

f dµ =

 A



u+ d µ −

A

u− d µ + i

 A

v+ d µ − i

 A

v− d µ

(A.4)

where f = u + iv and u+ , u− , v+ , and v− are the positive and negative parts of u and v, respectively. The restriction  X

| f | dµ < ∞

(A.5)

guarantees that each of the four integrals on the right-hand side of (A.4) is finite. Measurable complex-valued functions which satisfy (A.5) are said to be Lebesgue  integrable with respect to µ . Thus the notation X f d µ is used for nonnegative and for measurable, complexmeasurable functions, where X f d µ = ∞ is possible,  valued functions satisfying X | f | d µ < ∞ where X f d µ will be a (finite) value in C. Lebesgue integration is linear, meaning  X

(α f + g) d µ = α

 X

f dµ +

 X

g dµ

for α and complex measurable functions f and g satisfying  all complex scalars  | f | d µ < ∞ and |g| d µ < ∞. While this linearity is crucial, there are some subX X tleties in proving it. Theorem A.6 (Lebesgue Dominated Convergence Theorem). Suppose { fn } is a sequence of complex-valued µ -measurable functions and suppose lim fn (x) = f (x)

n→∞

for almost every x ∈ X. If there exists a measurable function g on X with | fn (x)| ≤ |g(x)| for almost every x ∈ X and

 X

|g|d µ < ∞

then f is measurable and 

lim

n→∞ X

fn d µ =

 X

f dµ.

As applications of the convergence theorems A.4 and A.6, and the linearity of the Lebesgue integral, we can give conditions under which term-by-term integration is allowed. This is an historically important issue, and with the Riemann integral instead of the Lebesgue integral sufficient conditions for such integration of series are often too restrictive. Theorem A.7. Suppose { fn } is a sequence of µ -measurable functions. (a)

If for all n, fn : X → [0, ∞], and we define

196

A Real Analysis Topics ∞

f (x) =

∑ fn (x),

n=1

then



f dµ =

X

(b)







n=1 X

fn d µ .

If the fn are complex-valued and ∞





n=1 X

| fn | d µ < ∞,

then the series ∑∞ n=1 f n converges almost everywhere, and     ∞

X



dµ =

fn

n=1





n=1 X

fn d µ .

Part (a) is proved by applying the monotone convergence theorem to the sequence of partial sums of ∑∞ n=1 f n . We allow the possibility that the positive term series ∑∞ n=1 f n (x) fails to converge for some x, in which case f (x) = ∞. Part (b) is proved by applying the dominated convergence theorem to the sequence of partial sums, which are dominated by g(x) ≡ ∑∞ n=1 | f n (x)|, where X g d µ < ∞ by (a).

A.3 L p Spaces For any measure space (X, M, µ ) and 1 ≤ p < ∞, set L p (X, µ ) ≡ {complex-valued measurable f :

 X

| f | p d µ < ∞}.

The notation L p (X, µ ) is often shortened to L p (X) or L p (µ ) when no confusion can result. We also define the space L∞ (X, µ ) of essentially bounded functions. We say that a measurable function is essentially bounded if there exists M < ∞ so that

µ ({x : | f (x)| > M}) = 0.

(A.6)

Equivalently, f is essentially bounded if it can be changed on a set of measure zero to produce a bounded function. When 1 ≤ p < ∞, L p (X, µ ) is a normed linear space if we define  f p ≡

 X

1 | f |pd µ

p

,

provided we agree to identify functions which agree µ -almost everywhere. If f ∈ L∞ (X, µ ), the infimum of all M > 0 that satisfy (A.6) is called the essential

A.4 The Stone–Weierstrass Theorem

197

supremum of f , and is denoted  f ∞ . This definition makes L∞ (X, µ ) a normed linear space. To verify the assertions of the last paragraph there are two issues to check: Is the sum of two L p functions still in L p , and does the triangle inequality hold for  ·  p ? For p = 1 and p = ∞, these hold as immediate consequences of the triangle inequality | f + g| ≤ | f | + |g| on C. When 1 < p < ∞ we need Minkowski’s inequality: For measurable functions f and g on X,  X

1 | f + g| p d µ

p



 X



1 | f |pd µ

p

+ X

1 |g| p d µ

p

.

Minkowski’s inequality follows from H¨older’s inequality  X

| f g|d µ ≤



| f | dµ p

X

 1  p X

1 |g| d µ q

q

(A.7)

where 1/p + 1/q = 1 (that is, p and q are conjugate indices.) H¨older’s inequality holds for the pair p = 1, q = ∞ if we replace the second integral on the right-hand side of (A.7) by g∞ . For the proofs of these two basic inequalities see Theorem 3.5 in [40], for example. By virtue of Example A.5, the space  p , as defined in Example 1.5 of Chapter 1, is the same as L p (N, µ ), where µ is counting measure on the subsets of N. This allows us to subsume the theory of  p into the theory of L p for general measure spaces. Since the only set with counting measure zero is the empty set, no “almost everywhere” conventions are needed with  p . The next result is fundamental. It asserts the completeness of the metric space L p (X, µ ). Theorem A.8 (Riesz–Fischer Theorem). For every positive measure µ and 1 ≤ p ≤ ∞, L p (µ ) is a Banach space. As discussed in Chapter 1, this theorem goes by the name of the Riesz–Fischer theorem, for simultaneous and independent work of Riesz and Fischer in the case p = 2. Fischer explicitly noted that Theorem A.8 requires “the use of notions of M. Lebesgue,” and that completeness does not hold if one considers continuous functions on [a, b] in the L2 (m) metric, since the L2 -limit of a sequence of continuous functions need not be continuous.

A.4 The Stone–Weierstrass Theorem Recall that for X a compact Hausdorff space, C(X) denotes the continuous complexvalued functions on X, endowed with the supremum norm  f ∞ = sup{| f (x)| : x ∈ X}.

198

A Real Analysis Topics

The real-valued functions in C(X) are denoted CR (X); this is a real vector space. The classical Weierstrass theorem says that for any finite interval [a, b] of the real line, the (real-valued) polynomials are dense in CR [a, b]; that is, given any f ∈ CR [a, b], there exist polynomials Pn converging uniformly to f on [a, b]. The same result holds for C[a, b], except now the polynomials are allowed to be complexvalued. This is a fundamental result in analysis, and it has numerous proofs. When the interval [a, b] is replaced by a compact Hausdorff space, it is not immediately clear how one might generalize Weierstrass’s result, as there is no notion of polynomials on general spaces. However, since the (real) polynomials on an interval are generated from sums, products, and real scalar products of the functions f (x) = 1 and g(x) = x, this suggests consideration of subalgebras of C(X). Recall that a (closed) subalgebra B of C(X) (or CR (X)) is a (closed) subspace that is closed under multiplication. We say that a subalgebra of C(X) separates points if given any x, y ∈ X there is an f ∈ B with f (x) = f (y). Marshall Stone generalized the Weierstrass theorem as follows: If B is a closed subalgebra of CR (X) that separates points and contains the constant functions, then B = CR (X). This is sometimes called the real Stone–Weierstrass theorem; a complexified version requires one additional hypothesis, namely that B be closed under conjugation (meaning that f ∈ B implies f ∈ B). Theorem A.9 (Stone–Weierstrass Theorem). If B is a closed subalgebra of C(X) that separates points, contains the constant functions, and is closed under conjugation, then B = C(X). To see why the extra hypothesis is needed, take X to be the closed unit disk D in C, and let B be the closed subalgebra of functions that are continuous in the closed disk and analytic in its interior, so that B separates points, contains the constants, but is not all of C(D), since, for example, f (z) = z does not belong to B. A proof of the Stone–Weierstrass Theorem can be found in [36].

A.5 Positive Linear Functionals on C(X) Let X be a compact Hausdorff space. The reader may find it convenient to think of X as a compact subset of the complex plane, since our principle application is to this setting. A positive linear functional on C(X) is a linear functional Λ : C(X) → C with the property that Λ ( f ) ≥ 0 if f ≥ 0 on X. We don’t a priori assume that Λ is bounded, but as a consequence of the positivity hypothesis, it will be so. By the Borel sets in X, we mean the smallest σ -algebra that contains all open sets in X. A measure defined on this σ -algebra is called a Borel measure, and a function which is measurable with respect to the Borel σ -algebra is called a Borel measurable function, or simply a Borel function. It should be clear that any function in C(X) is a Borel function. If we start with a finite, positive, Borel measure on X then C(X) ⊆ L1 (X, µ ), and the mapping

A.5 Positive Linear Functionals on C(X)

199

f→

 X

f dµ

will be a positive linear functional on C(X). The following result, known as the Riesz–Markov theorem, says that all positive linear functionals arise in this way. Theorem A.10 (Riesz–Markov Theorem). If X is a compact Hausdorff space and Λ is a positive linear functional on C(X), then there is a unique (positive) regular Borel measure µ on X with 

Λ( f) =

X

f dµ

for all f ∈ C(X). The adjective “regular” which appears in the statement will not be defined precisely here; see, for example [40]. Borel measures that are not regular are rather pathological. When X is a compact subset of the complex plane, for example, all Borel measures on X are regular. Notice that the norm of the linear functional Λ is µ (X).

References

1. D. Amir, Characterizations of Inner Product Spaces, Operator Theory: Advances and Applications, Vol. 20, Birkh¨auser, Basel, 1986. 2. W. Arveson, A Short Course on Spectral Theory, Springer-Verlag, New York, 2002. 3. B. Beauzamy, Introduction to Operator Theory and Invariant Subspaces, North-Holland, Amsterdam, 1988. 4. D. Bressoud, A Radical Approach to Lebesgue’s Theory of Integration, MAA Textbooks, Cambridge University Press, Cambridge, 2008. 5. S. Brown, Some invariant subspaces for subnormal operators, Integral Equations Operator Theory 1 (1978), 310–333. 6. N. Bourbaki, El´ements de Math´ematique: Espaces Vectoriels Topologiques Chap. I–V, Masson, Paris, 1981. 7. S.B. Chae, Lebesgue Integration, Second Edition, Springer-Verlag, New York, 1995. 8. J. Conway, A Course in Functional Analysis, Second Edition, Springer-Verlag, New York, 1990. 9. A. M. Davie, Review of “On the invariant subspace problem for Banach spaces” by P. Enflo, Mathematical Reviews on the Web, MR892591, 1988. http://www.ams.org/mathscinet 10. J. Dieudonn´e, History of Functional Analysis, Mathematics Studies 49, North-Holland, Amsterdam, 1981. 11. R. Douglas, Banach Algebra Techniques in Operator Theory, Pure and Applied Mathematics, Vol. 49. Academic Press, New York–London, 1972. 12. N. Dunford and J. Schwartz, Linear Operators, Part I: General Theory, Interscience, New York, 1958. 13. M. Fr´echet, Espaces abstraits, Gauthier-Villars, Paris, 1928. 14. C. Goffman and G. Pedrick, First Course in Functional Analysis, Prentice Hall, Englewood Cliffs, 1965. 15. P. Halmos, What does the spectral theorem say? Amer. Math. Monthly 70 (1963), 241–247. 16. P. Halmos, The legend of John von Neumann, Amer. Math. Monthly 80 (1973), 382–394. 17. P. Halmos, A Hilbert Space Problem Book, Second Edition, Springer-Verlag, New York, 1982. 18. P. Halmos, I Want to be a Mathematician: An Automathography, Springer-Verlag, New York, 1985. 19. D. Hilbert, Grundz¨uge einer allgemeinen Theorie der linearen Integralgleichungen, Chelsea, New York, 1953. 20. H. Hochstadt, Eduard Helly, father of the Hahn-Banach theorem, Math. Intelligencer 2 (1979/1980), No. 3, 123–125. 21. J. Horv´ath, On the Riesz–Fischer theorem, Studia Sci. Math. Hungar. 41 (2004), 467–478. ¨ 22. E. Helly, Uber Systeme linearer Gleichungen mit unendlich vielen Unbekannten, Monatsh. Math. Phys. 31 (1921), 60–91.

201

202

References

23. R.C. James, A non-reflexive Banach space isometric with its second conjugate space , Proc. Nat. Acad. Sci. U.S.A. 37 (1951), 174–177. 24. R.C. James, Characterizations of reflexivity, Studia Math. 23 (1963/1964), 205–216. 25. R. Kadison, Notes on the Gelfand–Neumark Theorem, in C∗ -Algebras: 1943–1993, A Fifty Year Celebration, Contemp. Math., 167, Amer. Math. Soc., Providence, RI 1994. 26. R. Kaluza, Through a reporter’s eyes: The Life of Stefan Banach, Birkh¨auser, Boston, 1996. 27. S. Krantz, Mathematical Aprochypha, Mathematical Association of America, Washington DC, 2002. 28. S. Krein and Yu. Petunin, Scales of Banach spaces, Russian Math. Surveys 21 (1966), 85–159. 29. H. Lebesgue, Measure and the Integral, Holden-Day, San Francisco, 1966. 30. N. Macrae, John von Neumann, Amer. Math. Soc., Providence, 1999. 31. R.D. Mauldin (Ed.), The Scottish Book: Mathematics from the Scottish Caf´e, Birkh¨auser, Boston, 1981. 32. V. Milman, Observations on the movement of people and ideas in twentieth-century mathematics, 215–242, in Mathematical Events of the Twentieth Century, A. Bolibruch, Yu. Osipov, YA. Sinai (Eds.), Springer-Verlag, Berlin, PHASIS, Moscow, 2006. 33. J. Munkres, Topology: A First Course, Prentice-Hall, Englewood Cliffs, NJ, 1975. 34. A. Pietsch, History of Banach Spaces and Linear Operators, Birkh¨auser, Boston, 2007. 35. H. Radjavi and P. Rosenthal, Invariant Subspaces, Second Edition, Dover Publications, Mineola, NY, 2003. 36. M. Reed and B. Simon, Methods of Modern Mathematical Physics I: Functional Analysis, Second Edition, Academic Press, New York, 1980. 37. M. R´edei, Ed., John von Neumann: Selected Letters, Amer. Math. Soc., London Math., Soc., Providence, 2005. 38. C. Reid, Hilbert, Springer-Verlag, Berlin, 1970. 39. H. Royden, Real Analysis, Third Edition, Macmillan, New York, 1988. 40. W. Rudin, Real and Complex Analysis, Third Edition, McGraw-Hill, New York, 1987. 41. D. Sarason, The multiplication theorem for Fredholm operators, Amer. Math. Monthly 94 (1987), 68–70. 42. D. Sarason, The exact answer to a question of Shields, Math. Intelligencer 12, (1990), 18–19. 43. K. Saxe, Beginning Functional Analysis, Springer, New York, 2002. 44. H. Steinhaus, Stefan Banach, Studia Math. (Ser. Specjalna) Zesyt 1 (1963), 7–15. 45. A. Vershik, The life and fate of functional analysis in the twentieth century, 437–447, in Mathematical Events of the Twentieth Century, A. Bolibruch, Yu. Osipov, YA. Sinai (Eds.), Springer-Verlag, Berlin, PHASIS, Moscow, 2006. 46. N. Wiener, I Am a Mathematician, Doubleday, New York, 1956. 47. L. Young, Mathematicians and Their Times, Mathematics Studies 48, North-Holland, Amsterdam, 1981. 48. N. Young, An Introduction to Hilbert Space, Cambridge University Press, Cambridge, 1988.

Index

A ≤ B, 142 A ⊥ B, A⊥ , 11 A∗ , adjoint, 36, 42 B(X, F ), 173 C(X), 2, 108, 197 C∗ (A), 135 C1 [0, 1], 23 Cϕ , 43 H 2 , Hβ2 , 4 L2 (T ), 19 La2 (D), 8 L p (X, µ ), L∞ (X, µ ), 3, 196 L p [0, 1], 3 Mϕ , 32 PM , 32 Tϕ , 32 W , Wiener algebra, 138 X/M, quotient space, 66 X ∗ , dual space, 17 ˆ Gelfand transform, 133 Γ (A), A, Γ (T ), compression spectrum, 115 Π , quotient map, 66

n Hn , 29 2β , 4  p , ∞ , 3 f dE, 170 ·, · , 6 graph(T ), 64 ker T , 39 ran T , 46 µh,g (S), 169 σ (T ), 96, 111 σap (T ), σ p (T ), 115 ∑ ⊕An , 45 ∑ ⊕Hn , 29 c0 , 58

m, Lebesgue measure, 190 r(A), 115 x∗∗ , 52 C, R, 2 D, 8 Mn (C), 149 N, 3 N0 , 4 Z, 19 A+ , 141 B(X,Y ), B(X), 32 K (X), 81 MA , 124 *-homomorphism, 120 adjoint Banach algebra element, 108 Banach space operator, 42 Hilbert space operator, 34 Alaoglu theorem, 130 algebra C∗ -algebra, 108 Banach, 38, 107 commutative, 107, 108 complex, 38, 107 unital, 38, 107 σ -algebra, 3, 187 almost everywhere, 193 annihilator, 70, 74 approximate eigenvalue, 97, 146 approximate point spectrum, 115, 118, 146 approximation problem, 84 Aronszajn, Nachman, 95 ascent, 105 Atkinson’s theorem, 144 axiom of choice, 50, 189

203

204 backward shift, 33, 36 Baire category theorem, 56 Banach algebra, 38, 107 Banach limit, 70 Banach space, 5 functional, 18, 74 Banach, Stefan, 6, 67 Banach–Alaoglu theorem, 130 Banach–Steinhaus theorem, 49, 57 basis Hamel, 19, 21, 27, 68, 77 orthonormal, 19, 20, 27 Schauder, 84 Beauzamy. B., 96 Bergman space, 8, 10, 18, 21 Bernstein, Allen, 95 Bessel sequence, 75 Bessel’s inequality, 20 Bohnenblust, H., 54 Borel function, 198 functional calculus, 166, 182 measure, 198 set, 198 Borel σ -algebra, 189 bounded linear functional, 16, 17, 25 linear operator, 31 sesquilinear form, 35, 45 bounded below, 41, 45, 97 Bourbaki, Nicolas, 16 C∗ -algebra, 108 singly generated, 135 C∗ -identity, 108 C-symmetric, 183 Calkin algebra, 123, 145 Carleson, Lennart, 59 carrier of spectral measure, 175 Cauchy integral formula, 9 Cauchy sequence, 5 Cauchy–Schwarz inequality, 7, 24 closed graph theorem, 49, 64 closed linear span, 29 cluster set, 153 compact operator, 80 set, 77 complete metric space, 5, 23 complex algebra, 38 complex homomorphism, 120 composition operator, 43, 73, 105 compression spectrum, 97, 115, 118, 147 cone, 142

Index conjugate index, 197 conjugate linear, 6 conjugation, 183 continuity in a metric space, 16, 129 in a topological space, 129 continuous functional calculus, 141 continuous spectrum, 97, 119, 120 contraction, 155 convergence of net, 129 weak, 28 convex set, 12 countably additive, 187 counting measure, 3, 188 cyclic operator, 47 vector, 47, 158, 183 Day, M.M., 18 dense range, 41, 45 descent, 100 diagonal operator, 33, 82, 103 diagonal trick, 81 Dieudonn´e, Jean, 3 dimension, 20 direct sum of Hilbert spaces, 29, 162 of operators, 45, 162 directed set, 129 Dirichlet kernel, 59 disjoint union, 163 dominated convergence theorem, 195 dual space, 16, 17, 50, 51 DuBois-Reymond, 59 eigenspace, 88 eigenvalue, 88, 118 eigenvector, 88 Einstein, Albert, 109 Enflo, Per, 84, 95 equivalent norms, 65, 78 essential norm, 123 range, 150, 166 spectral radius, 123 spectrum, 97, 123 supremum, 3, 197 essentially bounded, 196 invertible, 124 normal, 124 Euclidean norm, 2

Index fiber, 154 finite rank, 44, 81–83, 85, 103 finite-dimensional, 77, 79, 80 Fischer, Ernst, 22 forward shift, 33, 36, 81 Fourier coefficients, 21 Fourier series, 21, 59 frame, 76 Fr´echet, Maurice, 1, 2, 6, 15, 17 Fredholm alternative, 100 Fredholm operator, 143 Fredholm, Erik, 145 functional Banach space, 18, 74 functional calculus Borel, 166, 182 continuous, 141 Gelfand map, 133 transform, 133 Gelfand, I, 109, 113 Gelfand–Mazur theorem, 113, 124 Gelfand–Naimark theorem, 110 gliding hump, 71 G´odel, Kurt, 109 Gram–Schmidt process, 20, 21, 26 graph, 64 Hadamard, Jacques, 1, 15 Hahn–Banach theorem, 49 Halmos, Paul, 62, 95, 109 Hamel basis, 19, 21, 27, 68, 77 Hardy space, 4, 27 weighted, 4, 25 Hausdorff, Felix, 5 Heine–Borel theorem, 77 Helly, Eduard, 54, 57 Hermitian, 6 operator, 36 Hilbert space, 8 separable, 82 Hilbert, David, 8, 98 Hilbert–Schmidt operator, 85, 86, 103, 105 Hildebrandt, T., 84 Hilden, Hugh, 95 H¨older’s inequality, 9, 197 homomorphism, 120 *-homomorphism, 120 complex, 120 ideal, 81, 121 in B(H ), 150 maximal, 121 proper, 121

205 idempotent, 166 inner product, 6 on L2 (X, µ ), 7 inner product space, 12 integral Lebesgue, 3, 187, 193, 194 Riemann, 3, 187, 190 integral operator, 33, 37, 86 interior, 55 invariant subspace, 91, 185 invariant subspace problem, 94, 159 inverse mapping theorem, 62 inversion spectral mapping theorem, 115 invertible element in Banach algebra, 110 essentially, 124 operator, 39, 40, 62 involution, 38, 107 isometry, 17 isomorphism *-isomorphism, 120 Banach algebra, 113 Banach space, 52 Hilbert space, 39 James, R.C., 52 kernel integral operator, 33 linear map, 39 Krantz, Steve, 8 Lax, Peter, 109 Lebesgue dominated convergence theorem, 195 Lebesgue integral, 3, 187, 193, 194 Lebesgue measure, 3, 188 Lebesgue, Henri, 187 left inverse, 110, 147 Legendre polynomials, 26 linear combination, 19 linear functional, 15–17, 25 multiplicative, 124 positive, 69, 198 linear manifold, 11 linear map, 31 linear operator, 31 linearly independent, 19 Liouville theorem, 112 Lomonosov, Victor, 95 L¨owig, H., 54 Lw´ow, 67 matrix of an operator, 33

206 maximal ideal, 121 maximal ideal space, 124 Mazur, Stanislaw, 84, 113 mean-value property, 9 measurable function, 192, 194 set, 188 space, 167 measure, 188 Borel, 198 complex, 188 finite, 188 σ -finite, 188 Lebesgue, 3, 188 outer, 189 positive, 3, 188 regular, 199 signed, 188 space, 3, 187 spectral, 167 metric space, 5 metric topology, 5 Milman, V. D., 6 Minkowski’s inequality, 3, 23, 197 monotone convergence theorem, 194 multiplication operator, 32, 37, 150 multiplicative linear functional, 124 Murray, Francis, 54, 109 nearest point property, 12, 13 net, 129 nilpotent, 148 norm, 2 equivalent, 65, 78 essential, 123 Euclidean, 2 norm topology, 47 norm-attaining, 34, 44 norm-preserving, 50 normal element in C∗ -algebra, 110 operator, 38 normed linear space, 2 nowhere dense, 56 one-norm, 64 one-step extension, 52 open map, 61 open mapping theorem, 49, 61 operator bounded, 31 compact, 80 diagonal, 33, 82, 103 finite rank, 44, 81–83, 103

Index Fredholm, 143 Hermitian, 36 Hilbert–Schmidt, 85, 86, 103, 105 integral, 33, 37, 86 invertible, 39, 40, 62 multiplication, 32, 37, 150 norm-attaining, 34, 44 normal, 38 self-adjoint, 36, 38 unitary, 39, 44 orbit, 47, 158 ordering on self-adjoint elements, 142 orthogonal, 11 orthogonal projection, 15, 27, 32, 46, 155 orthonormal, 19 orthonormal basis, 19, 20, 27 Bergman space, 21 outer measure, 189 parallelogram equality, 11, 25 Parseval’s identity, 20 partial order, 49 Peano, Giuseppe, 2 point spectrum, 118, 120, 147 polarization identity, 12, 45 positive element in C∗ -algebra, 141, 154 linear functional, 69, 198 positive definite, 6 principle of uniform boundedness, 49, 56 product of operators, 32 product topology, 130 projection, 27, 32, 46, 155 projection theorem, 13, 18 proper ideal, 121 Pythagorean theorem, 11 quadratic form, 45 quasinilpotent, 148 quotient C∗ -algebra, 123 Banach algebra, 122, 148 map, 66, 74 norm, 73 space, 66, 73 radical, 133 radius of convergence, 4, 24 Read, C., 96 reducing subspace, 91 reflexive, 52, 71 regular spectral measure, 184 representation, 175 residual spectrum, 97, 119, 120, 150

Index reverse triangle inequality, 2, 23 Riemann integral, 3, 187, 190 Riesz representation theorem, 17, 35 Riesz, Frederic, 13, 17, 28, 43, 109 Riesz–Fischer sequence, 75 Riesz–Fischer theorem, 5, 22, 24, 197 Riesz–Markov theorem, 160, 199 right inverse, 110 Robinson, Abraham, 95 Schauder basis, 84 Schauder’s theorem, 85 Schauder, Juliusz, 43, 62 Schmidt, Erhard, 23 Scottish Caf´e, 67 Scottish Problem Book, 67, 84 self-adjoint element in C∗ -algebra, 110 operator, 36, 38 semi-simple, 133 separable Hilbert space, 82 separate points, 51 sesquilinear form, 35, 45 shift backward, 33, 36 forward, 33, 36, 81 similar, 96 simple function, 192 Smith, Kennan, 95 Sobczyk, A., 54 SOT, see strong operator topology span, 19 spectral mapping theorem, 114, 141, 147 spectral measure, 167 carrier, 175 support, 168, 176 spectral permanence, 116, 136 spectral radius, 115 essential, 123 formula, 116 spectral theorem, 157 for compact self-adjoint operators, 87 multiplication version, 158, 165 spectral measure version, 165, 181 spectrum of element in Banach algebra, 111 approximate point, 115, 118, 146 compression, 97, 115, 118, 147 continuous, 97, 119, 120 essential, 97, 123 of isometry, 148 of operator, 96

207 point, 118, 120, 147 residual, 97, 119, 120, 150 Steinhaus, Hugo, 6 Stone–Weierstrass theorem, 134, 198 strong operator topology, 47, 75, 168 strongly analytic, 112, 146 subadditive, 118 submultiplicative, 118 subnormal operator, 95 subspace, 6, 11 closed, 9, 11 support of spectral measure, 168, 176 supremum norm, 5 Toeplitz operator, 32, 37 topologies, comparing, 151 topology of pointwise convergence, 128 total order, 49 translation invariant, 189 triangle inequality, 2, 5 two-norm theorem, 65 Tychonoff theorem, 130, 131 uniform boundedness, see principle of uniform boundedness uniform operator topology, 47 unitarily equivalent, 44 unitarily equivalent to multiplication, 158 unitary element in C∗ -algebra, 146 operator, 39, 44 Volterra operator, 34, 104, 149, 183 von Neumann algebra, 109 von Neumann, John, 7, 85, 95, 109 weak convergence, 28 weak operator topology, 47 weak topology, 127 weak* topology, 127, 151 weakly analytic, 112, 146 Weierstrass theorem, 198 weighted Hardy space, 4, 25 weighted sequence space, 4, 5, 24 weighted shift, 44, 148 Weigl Institute, 68 Weigl, Rudolf Stefan, 68 Wiener algebra, 138, 152 Wiener, Norbert, 6 WOT, see weak operator topology Young, Laurence, 8 Zorn’s lemma, 19, 50, 68